Merge 04dd37f58f into 317218a682

2025-10-06 19:35:47 +00:00 · 2025-10-06 19:35:47 +00:00 · f9a61d9dd0
parent 317218a682 04dd37f58f
commit f9a61d9dd0
30 changed files with 3832 additions and 43 deletions
--- a/.devcontainer/install-deps.sh
+++ b/.devcontainer/install-deps.sh
@ -5,7 +5,7 @@ function sudo() {
  if [ "$UID" -eq 0 ]; then
    "$@"
  else
-    ${SUDO_PATH} "$@"
+    ${SUDO_PATH} -E "$@"
  fi
 }

@ -16,7 +16,7 @@ sudo apt-get update && \
 sudo apt-get install -y --no-install-recommends \
  build-essential \
  device-tree-compiler \
-  gperf g++-multilib gcc-multilib \
+  gperf \
  libnl-3-dev libdbus-1-dev libelf-dev libmpc-dev dwarves \
  bc openssl flex bison libssl-dev python3 python-is-python3 texinfo kmod cmake \
  wget zstd \
@ -30,6 +30,34 @@ pushd "${BUILDKIT_TMPDIR}" > /dev/null

 wget https://github.com/jetkvm/rv1106-system/releases/download/${BUILDKIT_VERSION}/buildkit.tar.zst && \
    sudo mkdir -p /opt/jetkvm-native-buildkit && \
-    sudo tar --use-compress-program="unzstd --long=31" -xvf buildkit.tar.zst -C /opt/jetkvm-native-buildkit && \
+    sudo tar --use-compress-program="zstd -d --long=31" -xvf buildkit.tar.zst -C /opt/jetkvm-native-buildkit && \
    rm buildkit.tar.zst
 popd
+
+# Install audio dependencies (ALSA and Opus) for JetKVM
+echo "Installing JetKVM audio dependencies..."
+SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
+PROJECT_ROOT="$(dirname "${SCRIPT_DIR}")"
+AUDIO_DEPS_SCRIPT="${PROJECT_ROOT}/install_audio_deps.sh"
+
+if [ -f "${AUDIO_DEPS_SCRIPT}" ]; then
+    echo "Running audio dependencies installation..."
+    # Pre-create audio libs directory with proper permissions
+    sudo mkdir -p /opt/jetkvm-audio-libs
+    sudo chmod 777 /opt/jetkvm-audio-libs
+    # Run installation script (now it can write without sudo)
+    bash "${AUDIO_DEPS_SCRIPT}"
+    echo "Audio dependencies installation completed."
+    if [ -d "/opt/jetkvm-audio-libs" ]; then
+        echo "Audio libraries installed in /opt/jetkvm-audio-libs"
+        # Set recursive permissions for all subdirectories and files
+        sudo chmod -R 777 /opt/jetkvm-audio-libs
+        echo "Permissions set to allow all users access to audio libraries"
+    else
+        echo "Error: /opt/jetkvm-audio-libs directory not found after installation."
+        exit 1
+    fi
+else
+    echo "Warning: Audio dependencies script not found at ${AUDIO_DEPS_SCRIPT}"
+    echo "Skipping audio dependencies installation."
+fi
--- a/.devcontainer/install_audio_deps.sh
+++ b/.devcontainer/install_audio_deps.sh
@ -0,0 +1,74 @@
+#!/bin/bash
+# .devcontainer/install_audio_deps.sh
+# Build ALSA and Opus static libs for ARM in /opt/jetkvm-audio-libs
+set -e
+
+# Sudo wrapper function
+SUDO_PATH=$(which sudo 2>/dev/null || echo "")
+function use_sudo() {
+  if [ "$UID" -eq 0 ]; then
+    "$@"
+  elif [ -n "$SUDO_PATH" ]; then
+    ${SUDO_PATH} -E "$@"
+  else
+    "$@"
+  fi
+}
+
+# Accept version parameters or use defaults
+ALSA_VERSION="${1:-1.2.14}"
+OPUS_VERSION="${2:-1.5.2}"
+
+AUDIO_LIBS_DIR="/opt/jetkvm-audio-libs"
+BUILDKIT_PATH="/opt/jetkvm-native-buildkit"
+BUILDKIT_FLAVOR="arm-rockchip830-linux-uclibcgnueabihf"
+CROSS_PREFIX="$BUILDKIT_PATH/bin/$BUILDKIT_FLAVOR"
+
+# Create directory with proper permissions
+use_sudo mkdir -p "$AUDIO_LIBS_DIR"
+use_sudo chmod 777 "$AUDIO_LIBS_DIR"
+cd "$AUDIO_LIBS_DIR"
+
+# Download sources
+[ -f alsa-lib-${ALSA_VERSION}.tar.bz2 ] || wget -N https://www.alsa-project.org/files/pub/lib/alsa-lib-${ALSA_VERSION}.tar.bz2
+[ -f opus-${OPUS_VERSION}.tar.gz ] || wget -N https://downloads.xiph.org/releases/opus/opus-${OPUS_VERSION}.tar.gz
+
+# Extract
+[ -d alsa-lib-${ALSA_VERSION} ] || tar xf alsa-lib-${ALSA_VERSION}.tar.bz2
+[ -d opus-${OPUS_VERSION} ] || tar xf opus-${OPUS_VERSION}.tar.gz
+
+# Optimization flags for ARM Cortex-A7 with NEON (simplified to avoid FD_SETSIZE issues)
+OPTIM_CFLAGS="-O2 -mfpu=neon -mtune=cortex-a7 -mfloat-abi=hard"
+
+export CC="${CROSS_PREFIX}-gcc"
+export CFLAGS="$OPTIM_CFLAGS"
+export CXXFLAGS="$OPTIM_CFLAGS"
+
+# Build ALSA
+cd alsa-lib-${ALSA_VERSION}
+if [ ! -f .built ]; then
+  chown -R $(whoami):$(whoami) .
+  # Use minimal ALSA configuration to avoid FD_SETSIZE issues in devcontainer
+  CFLAGS="$OPTIM_CFLAGS" ./configure --host $BUILDKIT_FLAVOR \
+    --enable-static=yes --enable-shared=no \
+    --with-pcm-plugins=rate,linear \
+    --disable-seq --disable-rawmidi --disable-ucm \
+    --disable-python --disable-old-symbols \
+    --disable-topology --disable-hwdep --disable-mixer \
+    --disable-alisp --disable-aload --disable-resmgr
+  make -j$(nproc)
+  touch .built
+fi
+cd ..
+
+# Build Opus
+cd opus-${OPUS_VERSION}
+if [ ! -f .built ]; then
+  chown -R $(whoami):$(whoami) .
+  CFLAGS="$OPTIM_CFLAGS" ./configure --host $BUILDKIT_FLAVOR --enable-static=yes --enable-shared=no --enable-fixed-point
+  make -j$(nproc)
+  touch .built
+fi
+cd ..
+
+echo "ALSA and Opus built in $AUDIO_LIBS_DIR"
--- a/.gitignore
+++ b/.gitignore
@ -13,3 +13,4 @@ node_modules
 # generated during the build process
 #internal/native/include
 #internal/native/lib
+internal/audio/bin/
--- a/Dockerfile.build
+++ b/Dockerfile.build
@ -6,6 +6,8 @@ ENV GOPATH=/go
 ENV PATH=$GOPATH/bin:/usr/local/go/bin:$PATH

 COPY install-deps.sh /install-deps.sh
+COPY install_audio_deps.sh /install_audio_deps.sh
+
 RUN /install-deps.sh

 # Create build directory
--- a/144
+++ b/144
@ -1,16 +1,59 @@
-BRANCH    := $(shell git rev-parse --abbrev-ref HEAD)
-BUILDDATE := $(shell date -u +%FT%T%z)
-BUILDTS   := $(shell date -u +%s)
-REVISION  := $(shell git rev-parse HEAD)
+# Build ALSA and Opus static libs for ARM in /opt/jetkvm-audio-libs
+build_audio_deps:
+	bash .devcontainer/install_audio_deps.sh $(ALSA_VERSION) $(OPUS_VERSION)
+
+# Prepare everything needed for local development (toolchain + audio deps + Go tools)
+dev_env: build_audio_deps
+	$(CLEAN_GO_CACHE)
+	@echo "Installing Go development tools..."
+	go install golang.org/x/tools/cmd/goimports@latest
+	@echo "Development environment ready."
+JETKVM_HOME ?= $(HOME)/.jetkvm
+BUILDKIT_PATH ?= /opt/jetkvm-native-buildkit
+BUILDKIT_FLAVOR ?= arm-rockchip830-linux-uclibcgnueabihf
+AUDIO_LIBS_DIR ?= /opt/jetkvm-audio-libs
+
+BRANCH    ?= $(shell git rev-parse --abbrev-ref HEAD)
+BUILDDATE ?= $(shell date -u +%FT%T%z)
+BUILDTS   ?= $(shell date -u +%s)
+REVISION  ?= $(shell git rev-parse HEAD)
 VERSION_DEV := 0.4.9-dev$(shell date +%Y%m%d%H%M)
 VERSION := 0.4.8

+
+# Audio library versions
+ALSA_VERSION ?= 1.2.14
+OPUS_VERSION ?= 1.5.2
+
+# Set PKG_CONFIG_PATH globally for all targets that use CGO with audio libraries
+export PKG_CONFIG_PATH := $(AUDIO_LIBS_DIR)/alsa-lib-$(ALSA_VERSION)/utils:$(AUDIO_LIBS_DIR)/opus-$(OPUS_VERSION)
+
+# Common command to clean Go cache with verbose output for all Go builds
+CLEAN_GO_CACHE := @echo "Cleaning Go cache..."; go clean -cache -v
+
+# Optimization flags for ARM Cortex-A7 with NEON SIMD
+OPTIM_CFLAGS := -O3 -mfpu=neon -mtune=cortex-a7 -mfloat-abi=hard -ftree-vectorize -ffast-math -funroll-loops -mvectorize-with-neon-quad -marm -D__ARM_NEON
+
+# Cross-compilation environment for ARM - exported globally
+export GOOS := linux
+export GOARCH := arm
+export GOARM := 7
+export CC := $(BUILDKIT_PATH)/bin/$(BUILDKIT_FLAVOR)-gcc
+export CGO_ENABLED := 1
+export CGO_CFLAGS := $(OPTIM_CFLAGS) -I$(BUILDKIT_PATH)/$(BUILDKIT_FLAVOR)/include -I$(BUILDKIT_PATH)/$(BUILDKIT_FLAVOR)/sysroot/usr/include
+export CGO_LDFLAGS := -L$(BUILDKIT_PATH)/$(BUILDKIT_FLAVOR)/lib -L$(BUILDKIT_PATH)/$(BUILDKIT_FLAVOR)/sysroot/usr/lib -lrockit -lrockchip_mpp -lrga -lpthread -lm -ldl
+
+# Audio-specific flags (only used for audio C binaries, NOT for main Go app)
+AUDIO_CFLAGS := $(CGO_CFLAGS) -I$(AUDIO_LIBS_DIR)/alsa-lib-$(ALSA_VERSION)/include -I$(AUDIO_LIBS_DIR)/opus-$(OPUS_VERSION)/include -I$(AUDIO_LIBS_DIR)/opus-$(OPUS_VERSION)/celt
+AUDIO_LDFLAGS := $(AUDIO_LIBS_DIR)/alsa-lib-$(ALSA_VERSION)/src/.libs/libasound.a $(AUDIO_LIBS_DIR)/opus-$(OPUS_VERSION)/.libs/libopus.a -lm -ldl -lpthread
+
 PROMETHEUS_TAG := github.com/prometheus/common/version
 KVM_PKG_NAME := github.com/jetkvm/kvm

 BUILDKIT_FLAVOR := arm-rockchip830-linux-uclibcgnueabihf
 BUILDKIT_PATH ?= /opt/jetkvm-native-buildkit
 SKIP_NATIVE_IF_EXISTS ?= 0
+SKIP_AUDIO_BINARIES_IF_EXISTS ?= 0
 SKIP_UI_BUILD ?= 0
 GO_BUILD_ARGS := -tags netgo,timetzdata,nomsgpack
 GO_RELEASE_BUILD_ARGS := -trimpath $(GO_BUILD_ARGS)
@ -49,22 +92,67 @@ build_native:
 			./scripts/build_cgo.sh; \
 	fi

-build_dev: build_native
+# Build audio output C binary (ALSA capture → Opus encode → IPC)
+build_audio_output: build_audio_deps
+	@if [ "$(SKIP_AUDIO_BINARIES_IF_EXISTS)" = "1" ] && [ -f "$(BIN_DIR)/jetkvm_audio_output" ]; then \
+		echo "jetkvm_audio_output already exists, skipping build..."; \
+	else \
+		echo "Building audio output binary (100% static)..."; \
+		mkdir -p $(BIN_DIR); \
+		$(CC) $(AUDIO_CFLAGS) -static \
+			-o $(BIN_DIR)/jetkvm_audio_output \
+			internal/audio/c/jetkvm_audio_output.c \
+			internal/audio/c/ipc_protocol.c \
+			internal/audio/c/audio_common.c \
+			internal/audio/c/audio.c \
+			$(AUDIO_LDFLAGS); \
+	fi
+
+# Build audio input C binary (IPC → Opus decode → ALSA playback)
+build_audio_input: build_audio_deps
+	@if [ "$(SKIP_AUDIO_BINARIES_IF_EXISTS)" = "1" ] && [ -f "$(BIN_DIR)/jetkvm_audio_input" ]; then \
+		echo "jetkvm_audio_input already exists, skipping build..."; \
+	else \
+		echo "Building audio input binary (100% static)..."; \
+		mkdir -p $(BIN_DIR); \
+		$(CC) $(AUDIO_CFLAGS) -static \
+			-o $(BIN_DIR)/jetkvm_audio_input \
+			internal/audio/c/jetkvm_audio_input.c \
+			internal/audio/c/ipc_protocol.c \
+			internal/audio/c/audio_common.c \
+			internal/audio/c/audio.c \
+			$(AUDIO_LDFLAGS); \
+	fi
+
+# Build both audio binaries and copy to embed location
+build_audio_binaries: build_audio_output build_audio_input
+	@echo "Audio binaries built successfully"
+	@echo "Copying binaries to embed location..."
+	@mkdir -p internal/audio/bin
+	@cp $(BIN_DIR)/jetkvm_audio_output internal/audio/bin/
+	@cp $(BIN_DIR)/jetkvm_audio_input internal/audio/bin/
+	@echo "Binaries ready for embedding"
+
+build_dev: build_native build_audio_deps build_audio_binaries
+	$(CLEAN_GO_CACHE)
 	@echo "Building..."
-	$(GO_CMD) build \
+	go build \
 		-ldflags="$(GO_LDFLAGS) -X $(KVM_PKG_NAME).builtAppVersion=$(VERSION_DEV)" \
 		$(GO_RELEASE_BUILD_ARGS) \
 		-o $(BIN_DIR)/jetkvm_app -v cmd/main.go

 build_test2json:
+	$(CLEAN_GO_CACHE)
 	$(GO_CMD) build -o $(BIN_DIR)/test2json cmd/test2json

 build_gotestsum:
+	$(CLEAN_GO_CACHE)
 	@echo "Building gotestsum..."
 	$(GO_CMD) install gotest.tools/gotestsum@latest
 	cp $(shell $(GO_CMD) env GOPATH)/bin/linux_arm/gotestsum $(BIN_DIR)/gotestsum

-build_dev_test: build_test2json build_gotestsum
+build_dev_test: build_audio_deps build_test2json build_gotestsum
+	$(CLEAN_GO_CACHE)
 # collect all directories that contain tests
 	@echo "Building tests for devices ..."
 	@rm -rf $(BIN_DIR)/tests && mkdir -p $(BIN_DIR)/tests
@ -74,7 +162,7 @@ build_dev_test: build_test2json build_gotestsum
 		test_pkg_name=$$(echo $$test | sed 's/^.\///g'); \
 		test_pkg_full_name=$(KVM_PKG_NAME)/$$(echo $$test | sed 's/^.\///g'); \
 		test_filename=$$(echo $$test_pkg_name | sed 's/\//__/g')_test; \
-		$(GO_CMD) test -v \
+		go test -v \
 			-ldflags="$(GO_LDFLAGS) -X $(KVM_PKG_NAME).builtAppVersion=$(VERSION_DEV)" \
 			$(GO_BUILD_ARGS) \
 			-c -o $(BIN_DIR)/tests/$$test_filename $$test; \
@ -111,9 +199,10 @@ dev_release: frontend build_dev
 	rclone copyto bin/jetkvm_app r2://jetkvm-update/app/$(VERSION_DEV)/jetkvm_app
 	rclone copyto bin/jetkvm_app.sha256 r2://jetkvm-update/app/$(VERSION_DEV)/jetkvm_app.sha256

-build_release: frontend build_native
+build_release: frontend build_native build_audio_deps build_audio_binaries
+	$(CLEAN_GO_CACHE)
 	@echo "Building release..."
-	$(GO_CMD) build \
+	go build \
 		-ldflags="$(GO_LDFLAGS) -X $(KVM_PKG_NAME).builtAppVersion=$(VERSION)" \
 		$(GO_RELEASE_BUILD_ARGS) \
 		-o bin/jetkvm_app cmd/main.go
@ -128,3 +217,38 @@ release:
 	@shasum -a 256 bin/jetkvm_app | cut -d ' ' -f 1 > bin/jetkvm_app.sha256
 	rclone copyto bin/jetkvm_app r2://jetkvm-update/app/$(VERSION)/jetkvm_app
 	rclone copyto bin/jetkvm_app.sha256 r2://jetkvm-update/app/$(VERSION)/jetkvm_app.sha256
+
+# Run both Go and UI linting
+lint: lint-go lint-ui
+	@echo "All linting completed successfully!"
+
+# Run golangci-lint locally with the same configuration as CI
+lint-go: build_audio_deps
+	@echo "Running golangci-lint..."
+	@mkdir -p static && touch static/.gitkeep
+	golangci-lint run --verbose
+
+# Run both Go and UI linting with auto-fix
+lint-fix: lint-go-fix lint-ui-fix
+	@echo "All linting with auto-fix completed successfully!"
+
+# Run golangci-lint with auto-fix
+lint-go-fix: build_audio_deps
+	@echo "Running golangci-lint with auto-fix..."
+	@mkdir -p static && touch static/.gitkeep
+	golangci-lint run --fix --verbose
+
+# Run UI linting locally (mirrors GitHub workflow ui-lint.yml)
+lint-ui:
+	@echo "Running UI lint..."
+	@cd ui && npm ci
+	@cd ui && npm run lint
+
+# Run UI linting with auto-fix
+lint-ui-fix:
+	@echo "Running UI lint with auto-fix..."
+	@cd ui && npm ci
+	@cd ui && npm run lint:fix
+
+# Legacy alias for UI linting (for backward compatibility)
+ui-lint: lint-ui
--- a/audio.go
+++ b/audio.go
@ -0,0 +1,351 @@
+package kvm
+
+import (
+	"io"
+	"sync"
+	"sync/atomic"
+
+	"github.com/jetkvm/kvm/internal/audio"
+	"github.com/jetkvm/kvm/internal/logging"
+	"github.com/pion/webrtc/v4"
+	"github.com/rs/zerolog"
+)
+
+const (
+	socketPathOutput = "/var/run/audio_output.sock"
+	socketPathInput  = "/var/run/audio_input.sock"
+)
+
+var (
+	audioMutex           sync.Mutex
+	outputSupervisor     *audio.Supervisor
+	inputSupervisor      *audio.Supervisor
+	outputClient         *audio.IPCClient
+	inputClient          *audio.IPCClient
+	outputRelay          *audio.OutputRelay
+	inputRelay           *audio.InputRelay
+	audioInitialized     bool
+	activeConnections    atomic.Int32
+	audioLogger          zerolog.Logger
+	currentAudioTrack    *webrtc.TrackLocalStaticSample
+	inputTrackHandling   atomic.Bool
+	useUSBForAudioOutput bool
+	audioOutputEnabled   atomic.Bool
+	audioInputEnabled    atomic.Bool
+)
+
+func initAudio() {
+	audioLogger = logging.GetDefaultLogger().With().Str("component", "audio-manager").Logger()
+
+	if err := audio.ExtractEmbeddedBinaries(); err != nil {
+		audioLogger.Error().Err(err).Msg("Failed to extract audio binaries")
+		return
+	}
+
+	// Load audio output source from config
+	ensureConfigLoaded()
+	useUSBForAudioOutput = config.AudioOutputSource == "usb"
+
+	// Enable both by default
+	audioOutputEnabled.Store(true)
+	audioInputEnabled.Store(true)
+
+	audioLogger.Debug().
+		Str("source", config.AudioOutputSource).
+		Msg("Audio subsystem initialized")
+	audioInitialized = true
+}
+
+// startAudioSubprocesses starts audio subprocesses and relays (skips already running ones)
+func startAudioSubprocesses() error {
+	audioMutex.Lock()
+	defer audioMutex.Unlock()
+
+	if !audioInitialized {
+		audioLogger.Warn().Msg("Audio not initialized, skipping subprocess start")
+		return nil
+	}
+
+	// Start output subprocess if not running and enabled
+	if outputSupervisor == nil && audioOutputEnabled.Load() {
+		alsaDevice := "hw:0,0" // HDMI
+		if useUSBForAudioOutput {
+			alsaDevice = "hw:1,0" // USB
+		}
+
+		outputSupervisor = audio.NewSupervisor(
+			"audio-output",
+			audio.GetAudioOutputBinaryPath(),
+			socketPathOutput,
+			[]string{
+				"ALSA_CAPTURE_DEVICE=" + alsaDevice,
+				"OPUS_BITRATE=128000",
+				"OPUS_COMPLEXITY=2",
+			},
+		)
+
+		if err := outputSupervisor.Start(); err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to start audio output supervisor")
+			outputSupervisor = nil
+			return err
+		}
+
+		outputClient = audio.NewIPCClient("audio-output", socketPathOutput, 0x4A4B4F55)
+
+		if currentAudioTrack != nil {
+			outputRelay = audio.NewOutputRelay(outputClient, currentAudioTrack)
+			if err := outputRelay.Start(); err != nil {
+				audioLogger.Error().Err(err).Msg("Failed to start audio output relay")
+			}
+		}
+	}
+
+	// Start input subprocess if not running, USB audio enabled, and input enabled
+	ensureConfigLoaded()
+	if inputSupervisor == nil && audioInputEnabled.Load() && config.UsbDevices != nil && config.UsbDevices.Audio {
+		inputSupervisor = audio.NewSupervisor(
+			"audio-input",
+			audio.GetAudioInputBinaryPath(),
+			socketPathInput,
+			[]string{
+				"ALSA_PLAYBACK_DEVICE=hw:1,0",
+				"OPUS_BITRATE=128000",
+			},
+		)
+
+		if err := inputSupervisor.Start(); err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to start input supervisor")
+			inputSupervisor = nil
+			return err
+		}
+
+		inputClient = audio.NewIPCClient("audio-input", socketPathInput, 0x4A4B4D49)
+		inputRelay = audio.NewInputRelay(inputClient)
+		if err := inputRelay.Start(); err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to start input relay")
+		}
+	}
+
+	return nil
+}
+
+// stopOutputSubprocessLocked stops output subprocess (assumes mutex is held)
+func stopOutputSubprocessLocked() {
+	if outputRelay != nil {
+		outputRelay.Stop()
+		outputRelay = nil
+	}
+	if outputClient != nil {
+		outputClient.Disconnect()
+		outputClient = nil
+	}
+	if outputSupervisor != nil {
+		outputSupervisor.Stop()
+		outputSupervisor = nil
+	}
+}
+
+// stopInputSubprocessLocked stops input subprocess (assumes mutex is held)
+func stopInputSubprocessLocked() {
+	if inputRelay != nil {
+		inputRelay.Stop()
+		inputRelay = nil
+	}
+	if inputClient != nil {
+		inputClient.Disconnect()
+		inputClient = nil
+	}
+	if inputSupervisor != nil {
+		inputSupervisor.Stop()
+		inputSupervisor = nil
+	}
+}
+
+// stopAudioSubprocessesLocked stops all audio subprocesses (assumes mutex is held)
+func stopAudioSubprocessesLocked() {
+	stopOutputSubprocessLocked()
+	stopInputSubprocessLocked()
+}
+
+// stopAudioSubprocesses stops all audio subprocesses
+func stopAudioSubprocesses() {
+	audioMutex.Lock()
+	defer audioMutex.Unlock()
+	stopAudioSubprocessesLocked()
+}
+
+func onWebRTCConnect() {
+	count := activeConnections.Add(1)
+	if count == 1 {
+		if err := startAudioSubprocesses(); err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to start audio subprocesses")
+		}
+	}
+}
+
+func onWebRTCDisconnect() {
+	count := activeConnections.Add(-1)
+	if count == 0 {
+		// Stop audio immediately to release HDMI audio device which shares hardware with video device
+		stopAudioSubprocesses()
+	}
+}
+
+func setAudioTrack(audioTrack *webrtc.TrackLocalStaticSample) {
+	audioMutex.Lock()
+	defer audioMutex.Unlock()
+
+	currentAudioTrack = audioTrack
+
+	if outputRelay != nil {
+		outputRelay.Stop()
+		outputRelay = nil
+	}
+
+	if outputClient != nil {
+		outputRelay = audio.NewOutputRelay(outputClient, audioTrack)
+		if err := outputRelay.Start(); err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to start output relay")
+		}
+	}
+}
+
+// SetAudioOutputSource switches between HDMI and USB audio output
+func SetAudioOutputSource(useUSB bool) error {
+	audioMutex.Lock()
+	defer audioMutex.Unlock()
+
+	if useUSBForAudioOutput == useUSB {
+		return nil
+	}
+
+	useUSBForAudioOutput = useUSB
+
+	ensureConfigLoaded()
+	if useUSB {
+		config.AudioOutputSource = "usb"
+	} else {
+		config.AudioOutputSource = "hdmi"
+	}
+	if err := SaveConfig(); err != nil {
+		audioLogger.Error().Err(err).Msg("Failed to save config")
+		return err
+	}
+
+	stopOutputSubprocessLocked()
+
+	// Restart if there are active connections
+	if activeConnections.Load() > 0 {
+		audioMutex.Unlock()
+		err := startAudioSubprocesses()
+		audioMutex.Lock()
+		if err != nil {
+			audioLogger.Error().Err(err).Msg("Failed to restart audio output")
+			return err
+		}
+	}
+
+	return nil
+}
+
+func setPendingInputTrack(track *webrtc.TrackRemote) {
+	audioMutex.Lock()
+	defer audioMutex.Unlock()
+
+	// Start input track handler only once per WebRTC session
+	if inputTrackHandling.CompareAndSwap(false, true) {
+		go handleInputTrackForSession(track)
+	}
+}
+
+// SetAudioOutputEnabled enables or disables audio output
+func SetAudioOutputEnabled(enabled bool) error {
+	if audioOutputEnabled.Swap(enabled) == enabled {
+		return nil // Already in desired state
+	}
+
+	if enabled {
+		if activeConnections.Load() > 0 {
+			return startAudioSubprocesses()
+		}
+	} else {
+		audioMutex.Lock()
+		stopOutputSubprocessLocked()
+		audioMutex.Unlock()
+	}
+
+	return nil
+}
+
+// SetAudioInputEnabled enables or disables audio input
+func SetAudioInputEnabled(enabled bool) error {
+	if audioInputEnabled.Swap(enabled) == enabled {
+		return nil // Already in desired state
+	}
+
+	if enabled {
+		if activeConnections.Load() > 0 {
+			return startAudioSubprocesses()
+		}
+	} else {
+		audioMutex.Lock()
+		stopInputSubprocessLocked()
+		audioMutex.Unlock()
+	}
+
+	return nil
+}
+
+// handleInputTrackForSession runs for the entire WebRTC session lifetime
+// It continuously reads from the track and sends to whatever relay is currently active
+func handleInputTrackForSession(track *webrtc.TrackRemote) {
+	defer inputTrackHandling.Store(false)
+
+	audioLogger.Debug().
+		Str("codec", track.Codec().MimeType).
+		Str("track_id", track.ID()).
+		Msg("starting session-lifetime track handler")
+
+	for {
+		// Read RTP packet (must always read to keep track alive)
+		rtpPacket, _, err := track.ReadRTP()
+		if err != nil {
+			if err == io.EOF {
+				audioLogger.Debug().Msg("audio track ended")
+				return
+			}
+			audioLogger.Warn().Err(err).Msg("failed to read RTP packet")
+			continue
+		}
+
+		// Extract Opus payload
+		opusData := rtpPacket.Payload
+		if len(opusData) == 0 {
+			continue
+		}
+
+		// Only send if input is enabled
+		if !audioInputEnabled.Load() {
+			continue // Drop frame but keep reading
+		}
+
+		// Get client in single mutex operation (hot path optimization)
+		audioMutex.Lock()
+		client := inputClient
+		audioMutex.Unlock()
+
+		if client == nil {
+			continue // No relay, drop frame but keep reading
+		}
+
+		if !client.IsConnected() {
+			if err := client.Connect(); err != nil {
+				continue
+			}
+		}
+
+		if err := client.WriteMessage(0, opusData); err != nil {
+			client.Disconnect()
+		}
+	}
+}
--- a/config.go
+++ b/config.go
@ -104,6 +104,7 @@ type Config struct {
 	UsbDevices           *usbgadget.Devices     `json:"usb_devices"`
 	NetworkConfig        *network.NetworkConfig `json:"network_config"`
 	DefaultLogLevel      string                 `json:"default_log_level"`
+	AudioOutputSource    string                 `json:"audio_output_source"` // "hdmi" or "usb"
 }

 func (c *Config) GetDisplayRotation() uint16 {
@ -160,8 +161,9 @@ var defaultConfig = &Config{
 		Keyboard:      true,
 		MassStorage:   true,
 	},
-	NetworkConfig:   &network.NetworkConfig{},
-	DefaultLogLevel: "INFO",
+	NetworkConfig:     &network.NetworkConfig{},
+	DefaultLogLevel:   "INFO",
+	AudioOutputSource: "hdmi",
 }

 var (
--- a/internal/audio/c/audio.c
+++ b/internal/audio/c/audio.c
@ -0,0 +1,789 @@
+/*
+ * JetKVM Audio Processing Module
+ *
+ * Bidirectional audio processing optimized for ARM NEON SIMD:
+ * - OUTPUT PATH: TC358743 HDMI audio → Client speakers
+ *   Pipeline: ALSA hw:0,0 capture → Opus encode (128kbps, FEC enabled)
+ *
+ * - INPUT PATH: Client microphone → Device speakers
+ *   Pipeline: Opus decode (with FEC) → ALSA hw:1,0 playback
+ *
+ * Key features:
+ * - ARM NEON SIMD optimization for all audio operations
+ * - Opus in-band FEC for packet loss resilience
+ * - S16_LE @ 48kHz stereo, 20ms frames (960 samples)
+ */
+
+#include <alsa/asoundlib.h>
+#include <opus.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sched.h>
+#include <time.h>
+#include <signal.h>
+
+// ARM NEON SIMD support (always available on JetKVM's ARM Cortex-A7)
+#include <arm_neon.h>
+
+// RV1106 (Cortex-A7) has 64-byte cache lines
+#define CACHE_LINE_SIZE 64
+#define SIMD_ALIGN __attribute__((aligned(16)))
+#define CACHE_ALIGN __attribute__((aligned(CACHE_LINE_SIZE)))
+#define SIMD_PREFETCH(addr, rw, locality) __builtin_prefetch(addr, rw, locality)
+
+// Compile-time trace logging - disabled for production (zero overhead)
+#define TRACE_LOG(...) ((void)0)
+
+// ALSA device handles
+static snd_pcm_t *pcm_capture_handle = NULL;  // OUTPUT: TC358743 HDMI audio → client
+static snd_pcm_t *pcm_playback_handle = NULL; // INPUT: Client microphone → device speakers
+
+// ALSA device names
+static const char *alsa_capture_device = NULL;
+static const char *alsa_playback_device = NULL;
+
+// Opus codec instances
+static OpusEncoder *encoder = NULL;
+static OpusDecoder *decoder = NULL;
+
+// Audio format (S16_LE @ 48kHz stereo)
+static uint32_t sample_rate = 48000;
+static uint8_t channels = 2;
+static uint16_t frame_size = 960;  // 20ms frames at 48kHz
+
+static uint32_t opus_bitrate = 128000;
+static uint8_t opus_complexity = 2;
+static uint16_t max_packet_size = 1500;
+
+// Opus encoder constants (hardcoded for production)
+#define OPUS_VBR 1                      // VBR enabled
+#define OPUS_VBR_CONSTRAINT 0           // Unconstrained VBR (better for low-volume signals)
+#define OPUS_SIGNAL_TYPE 3002           // OPUS_SIGNAL_MUSIC (better transient handling)
+#define OPUS_BANDWIDTH 1105             // OPUS_BANDWIDTH_FULLBAND (20kHz, enabled by 128kbps bitrate)
+#define OPUS_DTX 0                      // DTX disabled (prevents audio drops)
+#define OPUS_LSB_DEPTH 16               // 16-bit depth
+
+// ALSA retry configuration
+static uint32_t sleep_microseconds = 1000;
+static uint32_t sleep_milliseconds = 1;  // Precomputed: sleep_microseconds / 1000
+static uint8_t max_attempts_global = 5;
+static uint32_t max_backoff_us_global = 500000;
+
+int jetkvm_audio_capture_init();
+void jetkvm_audio_capture_close();
+int jetkvm_audio_read_encode(void *opus_buf);
+
+int jetkvm_audio_playback_init();
+void jetkvm_audio_playback_close();
+int jetkvm_audio_decode_write(void *opus_buf, int opus_size);
+
+void update_audio_constants(uint32_t bitrate, uint8_t complexity,
+                           uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                           uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff);
+void update_audio_decoder_constants(uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                                    uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff);
+int update_opus_encoder_params(uint32_t bitrate, uint8_t complexity);
+
+
+/**
+ * Sync encoder configuration from Go to C
+ */
+void update_audio_constants(uint32_t bitrate, uint8_t complexity,
+                           uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                           uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff) {
+    opus_bitrate = bitrate;
+    opus_complexity = complexity;
+    sample_rate = sr;
+    channels = ch;
+    frame_size = fs;
+    max_packet_size = max_pkt;
+    sleep_microseconds = sleep_us;
+    sleep_milliseconds = sleep_us / 1000;  // Precompute for snd_pcm_wait
+    max_attempts_global = max_attempts;
+    max_backoff_us_global = max_backoff;
+}
+
+/**
+ * Sync decoder configuration from Go to C (no encoder-only params)
+ */
+void update_audio_decoder_constants(uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                                    uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff) {
+    sample_rate = sr;
+    channels = ch;
+    frame_size = fs;
+    max_packet_size = max_pkt;
+    sleep_microseconds = sleep_us;
+    sleep_milliseconds = sleep_us / 1000;  // Precompute for snd_pcm_wait
+    max_attempts_global = max_attempts;
+    max_backoff_us_global = max_backoff;
+}
+
+/**
+ * Initialize ALSA device names from environment variables
+ * Must be called before jetkvm_audio_capture_init or jetkvm_audio_playback_init
+ */
+static void init_alsa_devices_from_env(void) {
+    if (alsa_capture_device == NULL) {
+        alsa_capture_device = getenv("ALSA_CAPTURE_DEVICE");
+        if (alsa_capture_device == NULL || alsa_capture_device[0] == '\0') {
+            alsa_capture_device = "hw:0,0"; // Default to HDMI
+        }
+    }
+    if (alsa_playback_device == NULL) {
+        alsa_playback_device = getenv("ALSA_PLAYBACK_DEVICE");
+        if (alsa_playback_device == NULL || alsa_playback_device[0] == '\0') {
+            alsa_playback_device = "hw:1,0"; // Default to USB gadget
+        }
+    }
+}
+
+// SIMD-OPTIMIZED BUFFER OPERATIONS (ARM NEON)
+
+/**
+ * Clear audio buffer using NEON (16 samples/iteration with 2x unrolling)
+ */
+static inline void simd_clear_samples_s16(short * __restrict__ buffer, uint32_t samples) {
+    const int16x8_t zero = vdupq_n_s16(0);
+    uint32_t i = 0;
+
+    // Process 16 samples at a time (2x unrolled for better pipeline utilization)
+    uint32_t simd_samples = samples & ~15U;
+    for (; i < simd_samples; i += 16) {
+        vst1q_s16(&buffer[i], zero);
+        vst1q_s16(&buffer[i + 8], zero);
+    }
+
+    // Handle remaining 8 samples
+    if (i + 8 <= samples) {
+        vst1q_s16(&buffer[i], zero);
+        i += 8;
+    }
+
+    // Scalar: remaining samples
+    for (; i < samples; i++) {
+        buffer[i] = 0;
+    }
+}
+
+// INITIALIZATION STATE TRACKING
+
+static volatile sig_atomic_t capture_initializing = 0;
+static volatile sig_atomic_t capture_initialized = 0;
+static volatile sig_atomic_t playback_initializing = 0;
+static volatile sig_atomic_t playback_initialized = 0;
+
+/**
+ * Update Opus encoder settings at runtime (does NOT modify FEC or hardcoded settings)
+ * @return 0 on success, -1 if not initialized, >0 if some settings failed
+ */
+int update_opus_encoder_params(uint32_t bitrate, uint8_t complexity) {
+    if (!encoder || !capture_initialized) {
+        return -1;
+    }
+
+    // Update runtime-configurable parameters
+    opus_bitrate = bitrate;
+    opus_complexity = complexity;
+
+    // Apply settings to encoder
+    int result = 0;
+    result |= opus_encoder_ctl(encoder, OPUS_SET_BITRATE(opus_bitrate));
+    result |= opus_encoder_ctl(encoder, OPUS_SET_COMPLEXITY(opus_complexity));
+
+    return result;
+}
+
+// ALSA UTILITY FUNCTIONS
+
+/**
+ * Open ALSA device with exponential backoff retry
+ * @return 0 on success, negative error code on failure
+ */
+// Helper: High-precision sleep using nanosleep (better than usleep)
+static inline void precise_sleep_us(uint32_t microseconds) {
+	struct timespec ts = {
+		.tv_sec = microseconds / 1000000,
+		.tv_nsec = (microseconds % 1000000) * 1000
+	};
+	nanosleep(&ts, NULL);
+}
+
+static int safe_alsa_open(snd_pcm_t **handle, const char *device, snd_pcm_stream_t stream) {
+	uint8_t attempt = 0;
+	int err;
+	uint32_t backoff_us = sleep_microseconds;
+
+	while (attempt < max_attempts_global) {
+		err = snd_pcm_open(handle, device, stream, SND_PCM_NONBLOCK);
+		if (err >= 0) {
+			snd_pcm_nonblock(*handle, 0);
+			return 0;
+		}
+
+		attempt++;
+
+		// Exponential backoff with bit shift (faster than multiplication)
+		if (err == -EBUSY || err == -EAGAIN) {
+			precise_sleep_us(backoff_us);
+			backoff_us = (backoff_us << 1 < max_backoff_us_global) ? (backoff_us << 1) : max_backoff_us_global;
+		} else if (err == -ENODEV || err == -ENOENT) {
+			precise_sleep_us(backoff_us << 1);
+			backoff_us = (backoff_us << 1 < max_backoff_us_global) ? (backoff_us << 1) : max_backoff_us_global;
+		} else if (err == -EPERM || err == -EACCES) {
+			precise_sleep_us(backoff_us >> 1);
+		} else {
+			precise_sleep_us(backoff_us);
+			backoff_us = (backoff_us << 1 < max_backoff_us_global) ? (backoff_us << 1) : max_backoff_us_global;
+		}
+	}
+	return err;
+}
+
+/**
+ * Configure ALSA device (S16_LE @ 48kHz stereo with optimized buffering)
+ * @param handle ALSA PCM handle
+ * @param device_name Unused (for debugging only)
+ * @return 0 on success, negative error code on failure
+ */
+static int configure_alsa_device(snd_pcm_t *handle, const char *device_name) {
+	snd_pcm_hw_params_t *params;
+	snd_pcm_sw_params_t *sw_params;
+	int err;
+
+	if (!handle) return -1;
+
+	snd_pcm_hw_params_alloca(&params);
+	snd_pcm_sw_params_alloca(&sw_params);
+
+	err = snd_pcm_hw_params_any(handle, params);
+	if (err < 0) return err;
+
+	err = snd_pcm_hw_params_set_access(handle, params, SND_PCM_ACCESS_RW_INTERLEAVED);
+	if (err < 0) return err;
+
+	err = snd_pcm_hw_params_set_format(handle, params, SND_PCM_FORMAT_S16_LE);
+	if (err < 0) return err;
+
+	err = snd_pcm_hw_params_set_channels(handle, params, channels);
+	if (err < 0) return err;
+
+	err = snd_pcm_hw_params_set_rate(handle, params, sample_rate, 0);
+	if (err < 0) {
+		unsigned int rate = sample_rate;
+		err = snd_pcm_hw_params_set_rate_near(handle, params, &rate, 0);
+		if (err < 0) return err;
+	}
+
+	snd_pcm_uframes_t period_size = frame_size;  // Optimized: use full frame as period
+	if (period_size < 64) period_size = 64;
+
+	err = snd_pcm_hw_params_set_period_size_near(handle, params, &period_size, 0);
+	if (err < 0) return err;
+
+	snd_pcm_uframes_t buffer_size = period_size * 2;  // Optimized: minimal buffer for low latency
+	err = snd_pcm_hw_params_set_buffer_size_near(handle, params, &buffer_size);
+	if (err < 0) return err;
+
+	err = snd_pcm_hw_params(handle, params);
+	if (err < 0) return err;
+
+	err = snd_pcm_sw_params_current(handle, sw_params);
+	if (err < 0) return err;
+
+	err = snd_pcm_sw_params_set_start_threshold(handle, sw_params, period_size);
+	if (err < 0) return err;
+
+	err = snd_pcm_sw_params_set_avail_min(handle, sw_params, period_size);
+	if (err < 0) return err;
+
+	err = snd_pcm_sw_params(handle, sw_params);
+	if (err < 0) return err;
+
+	return snd_pcm_prepare(handle);
+}
+
+// AUDIO OUTPUT PATH FUNCTIONS (TC358743 HDMI Audio → Client Speakers)
+
+/**
+ * Initialize OUTPUT path (TC358743 HDMI capture → Opus encoder)
+ * Opens hw:0,0 (TC358743) and creates Opus encoder with optimized settings
+ * @return 0 on success, -EBUSY if initializing, -1/-2/-3 on errors
+ */
+int jetkvm_audio_capture_init() {
+	int err;
+
+	init_alsa_devices_from_env();
+
+	if (__sync_bool_compare_and_swap(&capture_initializing, 0, 1) == 0) {
+		return -EBUSY;
+	}
+
+	if (capture_initialized) {
+		capture_initializing = 0;
+		return 0;
+	}
+
+	if (encoder) {
+		opus_encoder_destroy(encoder);
+		encoder = NULL;
+	}
+	if (pcm_capture_handle) {
+		snd_pcm_close(pcm_capture_handle);
+		pcm_capture_handle = NULL;
+	}
+
+	err = safe_alsa_open(&pcm_capture_handle, alsa_capture_device, SND_PCM_STREAM_CAPTURE);
+	if (err < 0) {
+		fprintf(stderr, "Failed to open ALSA capture device %s: %s\n",
+		        alsa_capture_device, snd_strerror(err));
+		fflush(stderr);
+		capture_initializing = 0;
+		return -1;
+	}
+
+	err = configure_alsa_device(pcm_capture_handle, "capture");
+	if (err < 0) {
+		snd_pcm_close(pcm_capture_handle);
+		pcm_capture_handle = NULL;
+		capture_initializing = 0;
+		return -2;
+	}
+
+	int opus_err = 0;
+	encoder = opus_encoder_create(sample_rate, channels, OPUS_APPLICATION_AUDIO, &opus_err);
+	if (!encoder || opus_err != OPUS_OK) {
+		if (pcm_capture_handle) {
+			snd_pcm_close(pcm_capture_handle);
+			pcm_capture_handle = NULL;
+		}
+		capture_initializing = 0;
+		return -3;
+	}
+
+	// Configure encoder with optimized settings
+	opus_encoder_ctl(encoder, OPUS_SET_BITRATE(opus_bitrate));
+	opus_encoder_ctl(encoder, OPUS_SET_COMPLEXITY(opus_complexity));
+	opus_encoder_ctl(encoder, OPUS_SET_VBR(OPUS_VBR));
+	opus_encoder_ctl(encoder, OPUS_SET_VBR_CONSTRAINT(OPUS_VBR_CONSTRAINT));
+	opus_encoder_ctl(encoder, OPUS_SET_SIGNAL(OPUS_SIGNAL_TYPE));
+	opus_encoder_ctl(encoder, OPUS_SET_BANDWIDTH(OPUS_BANDWIDTH));
+	opus_encoder_ctl(encoder, OPUS_SET_DTX(OPUS_DTX));
+	opus_encoder_ctl(encoder, OPUS_SET_LSB_DEPTH(OPUS_LSB_DEPTH));
+
+	opus_encoder_ctl(encoder, OPUS_SET_INBAND_FEC(1));
+	opus_encoder_ctl(encoder, OPUS_SET_PACKET_LOSS_PERC(20));
+
+	capture_initialized = 1;
+	capture_initializing = 0;
+	return 0;
+}
+
+/**
+ * Read HDMI audio, encode to Opus (OUTPUT path hot function)
+ * @param opus_buf Output buffer for encoded Opus packet
+ * @return >0 = Opus packet size in bytes, -1 = error
+ */
+__attribute__((hot)) int jetkvm_audio_read_encode(void * __restrict__ opus_buf) {
+	static short CACHE_ALIGN pcm_buffer[960 * 2];  // Cache-aligned
+	unsigned char * __restrict__ out = (unsigned char*)opus_buf;
+	int32_t pcm_rc, nb_bytes;
+	int32_t err = 0;
+	uint8_t recovery_attempts = 0;
+	const uint8_t max_recovery_attempts = 3;
+
+	// Prefetch for write (out) and read (pcm_buffer) - RV1106 has small L1 cache
+	SIMD_PREFETCH(out, 1, 0);  // Write, immediate use
+	SIMD_PREFETCH(pcm_buffer, 0, 0);  // Read, immediate use
+	SIMD_PREFETCH(pcm_buffer + 64, 0, 1);  // Prefetch next cache line
+
+	if (__builtin_expect(!capture_initialized || !pcm_capture_handle || !encoder || !opus_buf, 0)) {
+		TRACE_LOG("[AUDIO_OUTPUT] jetkvm_audio_read_encode: Failed safety checks - capture_initialized=%d, pcm_capture_handle=%p, encoder=%p, opus_buf=%p\n",
+				capture_initialized, pcm_capture_handle, encoder, opus_buf);
+		return -1;
+	}
+
+retry_read:
+	// Read 960 frames (20ms) from ALSA capture device
+	pcm_rc = snd_pcm_readi(pcm_capture_handle, pcm_buffer, frame_size);
+
+	if (__builtin_expect(pcm_rc < 0, 0)) {
+		if (pcm_rc == -EPIPE) {
+			recovery_attempts++;
+			if (recovery_attempts > max_recovery_attempts) {
+				return -1;
+			}
+			err = snd_pcm_prepare(pcm_capture_handle);
+			if (err < 0) {
+				snd_pcm_drop(pcm_capture_handle);
+				err = snd_pcm_prepare(pcm_capture_handle);
+				if (err < 0) return -1;
+			}
+			goto retry_read;
+		} else if (pcm_rc == -EAGAIN) {
+			// Wait for data to be available
+			snd_pcm_wait(pcm_capture_handle, sleep_milliseconds);
+			goto retry_read;
+		} else if (pcm_rc == -ESTRPIPE) {
+			recovery_attempts++;
+			if (recovery_attempts > max_recovery_attempts) {
+				return -1;
+			}
+			uint8_t resume_attempts = 0;
+			while ((err = snd_pcm_resume(pcm_capture_handle)) == -EAGAIN && resume_attempts < 10) {
+				snd_pcm_wait(pcm_capture_handle, sleep_milliseconds);
+				resume_attempts++;
+			}
+			if (err < 0) {
+				err = snd_pcm_prepare(pcm_capture_handle);
+				if (err < 0) return -1;
+			}
+			return 0;
+		} else if (pcm_rc == -ENODEV) {
+			return -1;
+		} else if (pcm_rc == -EIO) {
+			recovery_attempts++;
+			if (recovery_attempts <= max_recovery_attempts) {
+				snd_pcm_drop(pcm_capture_handle);
+				err = snd_pcm_prepare(pcm_capture_handle);
+				if (err >= 0) {
+					goto retry_read;
+				}
+			}
+			return -1;
+		} else {
+			recovery_attempts++;
+			if (recovery_attempts <= 1 && pcm_rc == -EINTR) {
+				goto retry_read;
+			} else if (recovery_attempts <= 1 && pcm_rc == -EBUSY) {
+				snd_pcm_wait(pcm_capture_handle, 1); // Wait 1ms for device
+				goto retry_read;
+			}
+			return -1;
+		}
+	}
+
+	// Zero-pad if we got a short read
+	if (__builtin_expect(pcm_rc < frame_size, 0)) {
+		uint32_t remaining_samples = (frame_size - pcm_rc) * channels;
+		simd_clear_samples_s16(&pcm_buffer[pcm_rc * channels], remaining_samples);
+	}
+
+	// Find peak amplitude with NEON SIMD
+	uint32_t total_samples = frame_size * channels;
+	int16x8_t vmax = vdupq_n_s16(0);
+
+	uint32_t i;
+	for (i = 0; i + 8 <= total_samples; i += 8) {
+		int16x8_t v = vld1q_s16(&pcm_buffer[i]);
+		int16x8_t vabs = vabsq_s16(v);
+		vmax = vmaxq_s16(vmax, vabs);
+	}
+
+	// Horizontal max reduction (manual for ARMv7)
+	int16x4_t vmax_low = vget_low_s16(vmax);
+	int16x4_t vmax_high = vget_high_s16(vmax);
+	int16x4_t vmax_reduced = vmax_s16(vmax_low, vmax_high);
+	vmax_reduced = vpmax_s16(vmax_reduced, vmax_reduced);
+	vmax_reduced = vpmax_s16(vmax_reduced, vmax_reduced);
+	int16_t peak = vget_lane_s16(vmax_reduced, 0);
+
+	// Handle remaining samples
+	for (; i < total_samples; i++) {
+		int16_t abs_val = (pcm_buffer[i] < 0) ? -pcm_buffer[i] : pcm_buffer[i];
+		if (abs_val > peak) peak = abs_val;
+	}
+
+	// Apply gain if signal is weak (below -18dB = 4096) but above noise floor
+	// Noise gate: only apply gain if peak > 256 (below this is likely just noise)
+	// Target: boost to ~50% of range (16384) to improve SNR
+	if (peak > 256 && peak < 4096) {
+		float gain = 16384.0f / peak;
+		if (gain > 8.0f) gain = 8.0f;  // Max 18dB boost
+
+		// Apply gain with NEON and saturation
+		float32x4_t vgain = vdupq_n_f32(gain);
+		for (i = 0; i + 8 <= total_samples; i += 8) {
+			int16x8_t v = vld1q_s16(&pcm_buffer[i]);
+
+			// Convert to float, apply gain, saturate back to int16
+			int32x4_t v_low = vmovl_s16(vget_low_s16(v));
+			int32x4_t v_high = vmovl_s16(vget_high_s16(v));
+
+			float32x4_t f_low = vcvtq_f32_s32(v_low);
+			float32x4_t f_high = vcvtq_f32_s32(v_high);
+
+			f_low = vmulq_f32(f_low, vgain);
+			f_high = vmulq_f32(f_high, vgain);
+
+			v_low = vcvtq_s32_f32(f_low);
+			v_high = vcvtq_s32_f32(f_high);
+
+			// Saturate to int16 range
+			int16x4_t result_low = vqmovn_s32(v_low);
+			int16x4_t result_high = vqmovn_s32(v_high);
+
+			vst1q_s16(&pcm_buffer[i], vcombine_s16(result_low, result_high));
+		}
+
+		// Handle remaining samples
+		for (; i < total_samples; i++) {
+			int32_t boosted = (int32_t)(pcm_buffer[i] * gain);
+			if (boosted > 32767) boosted = 32767;
+			if (boosted < -32768) boosted = -32768;
+			pcm_buffer[i] = (int16_t)boosted;
+		}
+	}
+
+	nb_bytes = opus_encode(encoder, pcm_buffer, frame_size, out, max_packet_size);
+	return nb_bytes;
+}
+
+// AUDIO INPUT PATH FUNCTIONS (Client Microphone → Device Speakers)
+
+/**
+ * Initialize INPUT path (Opus decoder → device speakers)
+ * Opens hw:1,0 (USB gadget) or "default" and creates Opus decoder
+ * @return 0 on success, -EBUSY if initializing, -1/-2 on errors
+ */
+int jetkvm_audio_playback_init() {
+	int err;
+
+	init_alsa_devices_from_env();
+
+	if (__sync_bool_compare_and_swap(&playback_initializing, 0, 1) == 0) {
+		return -EBUSY;
+	}
+
+	if (playback_initialized) {
+		playback_initializing = 0;
+		return 0;
+	}
+
+	if (decoder) {
+		opus_decoder_destroy(decoder);
+		decoder = NULL;
+	}
+	if (pcm_playback_handle) {
+		snd_pcm_close(pcm_playback_handle);
+		pcm_playback_handle = NULL;
+	}
+
+	err = safe_alsa_open(&pcm_playback_handle, alsa_playback_device, SND_PCM_STREAM_PLAYBACK);
+	if (err < 0) {
+		fprintf(stderr, "Failed to open ALSA playback device %s: %s\n",
+		        alsa_playback_device, snd_strerror(err));
+		fflush(stderr);
+		err = safe_alsa_open(&pcm_playback_handle, "default", SND_PCM_STREAM_PLAYBACK);
+		if (err < 0) {
+			playback_initializing = 0;
+			return -1;
+		}
+	}
+
+	err = configure_alsa_device(pcm_playback_handle, "playback");
+	if (err < 0) {
+		snd_pcm_close(pcm_playback_handle);
+		pcm_playback_handle = NULL;
+		playback_initializing = 0;
+		return -1;
+	}
+
+	int opus_err = 0;
+	decoder = opus_decoder_create(sample_rate, channels, &opus_err);
+	if (!decoder || opus_err != OPUS_OK) {
+		snd_pcm_close(pcm_playback_handle);
+		pcm_playback_handle = NULL;
+		playback_initializing = 0;
+		return -2;
+	}
+
+	playback_initialized = 1;
+	playback_initializing = 0;
+	return 0;
+}
+
+/**
+ * Decode Opus, write to device speakers (INPUT path hot function)
+ * Processing pipeline: Opus decode (with FEC) → ALSA playback with error recovery
+ * @param opus_buf Encoded Opus packet from client
+ * @param opus_size Size of Opus packet in bytes
+ * @return >0 = PCM frames written, 0 = frame skipped, -1/-2 = error
+ */
+__attribute__((hot)) int jetkvm_audio_decode_write(void * __restrict__ opus_buf, int32_t opus_size) {
+	static short CACHE_ALIGN pcm_buffer[960 * 2];  // Cache-aligned
+	unsigned char * __restrict__ in = (unsigned char*)opus_buf;
+	int32_t pcm_frames, pcm_rc, err = 0;
+	uint8_t recovery_attempts = 0;
+	const uint8_t max_recovery_attempts = 3;
+
+	// Prefetch input buffer - locality 0 for immediate use
+	SIMD_PREFETCH(in, 0, 0);
+
+	if (__builtin_expect(!playback_initialized || !pcm_playback_handle || !decoder || !opus_buf || opus_size <= 0, 0)) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Failed safety checks - playback_initialized=%d, pcm_playback_handle=%p, decoder=%p, opus_buf=%p, opus_size=%d\n",
+				playback_initialized, pcm_playback_handle, decoder, opus_buf, opus_size);
+		return -1;
+	}
+
+	if (opus_size > max_packet_size) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Opus packet too large - size=%d, max=%d\n", opus_size, max_packet_size);
+		return -1;
+	}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Processing Opus packet - size=%d bytes\n", opus_size);
+
+	// Decode Opus packet to PCM (FEC automatically applied if embedded in packet)
+	// decode_fec=0 means normal decode (FEC data is used automatically when present)
+	pcm_frames = opus_decode(decoder, in, opus_size, pcm_buffer, frame_size, 0);
+
+	if (__builtin_expect(pcm_frames < 0, 0)) {
+		// Decode failed - attempt packet loss concealment using FEC from previous packet
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Opus decode failed with error %d, attempting packet loss concealment\n", pcm_frames);
+
+		// decode_fec=1 means use FEC data from the NEXT packet to reconstruct THIS lost packet
+		pcm_frames = opus_decode(decoder, NULL, 0, pcm_buffer, frame_size, 1);
+		if (pcm_frames < 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Packet loss concealment also failed with error %d\n", pcm_frames);
+			return -1;
+		}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Packet loss concealment succeeded, recovered %d frames\n", pcm_frames);
+	} else
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Opus decode successful - decoded %d PCM frames\n", pcm_frames);
+
+retry_write:
+	// Write decoded PCM to ALSA playback device
+	pcm_rc = snd_pcm_writei(pcm_playback_handle, pcm_buffer, pcm_frames);
+	if (__builtin_expect(pcm_rc < 0, 0)) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: ALSA write failed with error %d (%s), attempt %d/%d\n",
+				pcm_rc, snd_strerror(pcm_rc), recovery_attempts + 1, max_recovery_attempts);
+
+		if (pcm_rc == -EPIPE) {
+			recovery_attempts++;
+			if (recovery_attempts > max_recovery_attempts) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Buffer underrun recovery failed after %d attempts\n", max_recovery_attempts);
+				return -2;
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Buffer underrun detected, attempting recovery (attempt %d)\n", recovery_attempts);
+			err = snd_pcm_prepare(pcm_playback_handle);
+			if (err < 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: snd_pcm_prepare failed (%s), trying drop+prepare\n", snd_strerror(err));
+				snd_pcm_drop(pcm_playback_handle);
+				err = snd_pcm_prepare(pcm_playback_handle);
+				if (err < 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: drop+prepare recovery failed (%s)\n", snd_strerror(err));
+					return -2;
+				}
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Buffer underrun recovery successful, retrying write\n");
+			goto retry_write;
+		} else if (pcm_rc == -ESTRPIPE) {
+			recovery_attempts++;
+			if (recovery_attempts > max_recovery_attempts) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device suspend recovery failed after %d attempts\n", max_recovery_attempts);
+				return -2;
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device suspended, attempting resume (attempt %d)\n", recovery_attempts);
+			uint8_t resume_attempts = 0;
+			while ((err = snd_pcm_resume(pcm_playback_handle)) == -EAGAIN && resume_attempts < 10) {
+				snd_pcm_wait(pcm_playback_handle, sleep_milliseconds);
+				resume_attempts++;
+			}
+			if (err < 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device resume failed (%s), trying prepare fallback\n", snd_strerror(err));
+				err = snd_pcm_prepare(pcm_playback_handle);
+				if (err < 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Prepare fallback failed (%s)\n", snd_strerror(err));
+					return -2;
+				}
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device suspend recovery successful, skipping frame\n");
+			return 0;
+		} else if (pcm_rc == -ENODEV) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device disconnected (ENODEV) - critical error\n");
+			return -2;
+		} else if (pcm_rc == -EIO) {
+			recovery_attempts++;
+			if (recovery_attempts <= max_recovery_attempts) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: I/O error detected, attempting recovery\n");
+				snd_pcm_drop(pcm_playback_handle);
+				err = snd_pcm_prepare(pcm_playback_handle);
+				if (err >= 0) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: I/O error recovery successful, retrying write\n");
+					goto retry_write;
+				}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: I/O error recovery failed (%s)\n", snd_strerror(err));
+			}
+			return -2;
+		} else if (pcm_rc == -EAGAIN) {
+			recovery_attempts++;
+			if (recovery_attempts <= max_recovery_attempts) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device not ready (EAGAIN), waiting and retrying\n");
+				snd_pcm_wait(pcm_playback_handle, 1);  // Wait 1ms
+				goto retry_write;
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Device not ready recovery failed after %d attempts\n", max_recovery_attempts);
+			return -2;
+		} else {
+			recovery_attempts++;
+			if (recovery_attempts <= 1 && (pcm_rc == -EINTR || pcm_rc == -EBUSY)) {
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Transient error %d (%s), retrying once\n", pcm_rc, snd_strerror(pcm_rc));
+				snd_pcm_wait(pcm_playback_handle, 1);  // Wait 1ms
+				goto retry_write;
+			}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Unrecoverable error %d (%s)\n", pcm_rc, snd_strerror(pcm_rc));
+			return -2;
+		}
+	}
+		TRACE_LOG("[AUDIO_INPUT] jetkvm_audio_decode_write: Successfully wrote %d PCM frames to device\n", pcm_frames);
+	return pcm_frames;
+}
+
+// CLEANUP FUNCTIONS
+
+/**
+ * Close INPUT path (thread-safe with drain)
+ */
+void jetkvm_audio_playback_close() {
+	while (playback_initializing) {
+		sched_yield();
+	}
+
+	if (__sync_bool_compare_and_swap(&playback_initialized, 1, 0) == 0) {
+		return;
+	}
+
+	if (decoder) {
+		opus_decoder_destroy(decoder);
+		decoder = NULL;
+	}
+	if (pcm_playback_handle) {
+		snd_pcm_drain(pcm_playback_handle);
+		snd_pcm_close(pcm_playback_handle);
+		pcm_playback_handle = NULL;
+	}
+}
+
+/**
+ * Close OUTPUT path (thread-safe with drain)
+ */
+void jetkvm_audio_capture_close() {
+	while (capture_initializing) {
+		sched_yield();
+	}
+
+	if (__sync_bool_compare_and_swap(&capture_initialized, 1, 0) == 0) {
+		return;
+	}
+
+	if (encoder) {
+		opus_encoder_destroy(encoder);
+		encoder = NULL;
+	}
+	if (pcm_capture_handle) {
+		snd_pcm_drain(pcm_capture_handle);
+		snd_pcm_close(pcm_capture_handle);
+		pcm_capture_handle = NULL;
+	}
+}
--- a/internal/audio/c/audio_common.c
+++ b/internal/audio/c/audio_common.c
@ -0,0 +1,169 @@
+/*
+ * JetKVM Audio Common Utilities
+ *
+ * Shared functions used by both audio input and output servers
+ */
+
+#include "audio_common.h"
+#include "ipc_protocol.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <signal.h>
+#include <unistd.h>
+#include <sys/socket.h>
+#include <time.h>
+
+// Forward declarations for encoder update (only in output server)
+extern int update_opus_encoder_params(uint32_t bitrate, uint8_t complexity);
+
+// GLOBAL STATE FOR SIGNAL HANDLER
+
+// Pointer to the running flag that will be set to 0 on shutdown
+static volatile sig_atomic_t *g_running_ptr = NULL;
+
+// SIGNAL HANDLERS
+
+static void signal_handler(int signo) {
+    if (signo == SIGTERM || signo == SIGINT) {
+        printf("Audio server: Received signal %d, shutting down...\n", signo);
+        if (g_running_ptr != NULL) {
+            *g_running_ptr = 0;
+        }
+    }
+}
+
+void audio_common_setup_signal_handlers(volatile sig_atomic_t *running) {
+    g_running_ptr = running;
+
+    struct sigaction sa;
+    memset(&sa, 0, sizeof(sa));
+    sa.sa_handler = signal_handler;
+    sigemptyset(&sa.sa_mask);
+    sa.sa_flags = 0;
+
+    sigaction(SIGTERM, &sa, NULL);
+    sigaction(SIGINT, &sa, NULL);
+
+    // Ignore SIGPIPE (write to closed socket should return error, not crash)
+    signal(SIGPIPE, SIG_IGN);
+}
+
+
+int32_t audio_common_parse_env_int(const char *name, int32_t default_value) {
+    const char *str = getenv(name);
+    if (str == NULL || str[0] == '\0') {
+        return default_value;
+    }
+    return (int32_t)atoi(str);
+}
+
+const char* audio_common_parse_env_string(const char *name, const char *default_value) {
+    const char *str = getenv(name);
+    if (str == NULL || str[0] == '\0') {
+        return default_value;
+    }
+    return str;
+}
+
+// COMMON CONFIGURATION
+
+void audio_common_load_config(audio_config_t *config, int is_output) {
+    // ALSA device configuration
+    if (is_output) {
+        config->alsa_device = audio_common_parse_env_string("ALSA_CAPTURE_DEVICE", "hw:0,0");
+    } else {
+        config->alsa_device = audio_common_parse_env_string("ALSA_PLAYBACK_DEVICE", "hw:1,0");
+    }
+
+    // Common Opus configuration
+    config->opus_bitrate = audio_common_parse_env_int("OPUS_BITRATE", 128000);
+    config->opus_complexity = audio_common_parse_env_int("OPUS_COMPLEXITY", 2);
+
+    // Audio format
+    config->sample_rate = audio_common_parse_env_int("AUDIO_SAMPLE_RATE", 48000);
+    config->channels = audio_common_parse_env_int("AUDIO_CHANNELS", 2);
+    config->frame_size = audio_common_parse_env_int("AUDIO_FRAME_SIZE", 960);
+
+    // Log configuration
+    printf("Audio %s Server Configuration:\n", is_output ? "Output" : "Input");
+    printf("  ALSA Device:    %s\n", config->alsa_device);
+    printf("  Sample Rate:    %d Hz\n", config->sample_rate);
+    printf("  Channels:       %d\n", config->channels);
+    printf("  Frame Size:     %d samples\n", config->frame_size);
+    if (is_output) {
+        printf("  Opus Bitrate:   %d bps\n", config->opus_bitrate);
+        printf("  Opus Complexity: %d\n", config->opus_complexity);
+    }
+}
+
+void audio_common_print_startup(const char *server_name) {
+    printf("JetKVM %s Starting...\n", server_name);
+}
+
+void audio_common_print_shutdown(const char *server_name) {
+    printf("Shutting down %s...\n", server_name);
+}
+
+
+int audio_common_handle_opus_config(const uint8_t *data, uint32_t length, int is_encoder) {
+    ipc_opus_config_t config;
+
+    if (ipc_parse_opus_config(data, length, &config) != 0) {
+        fprintf(stderr, "Failed to parse Opus config\n");
+        return -1;
+    }
+
+    if (is_encoder) {
+        printf("Received Opus config: bitrate=%u, complexity=%u\n",
+               config.bitrate, config.complexity);
+
+        int result = update_opus_encoder_params(
+            config.bitrate,
+            config.complexity
+        );
+
+        if (result != 0) {
+            fprintf(stderr, "Warning: Failed to apply Opus encoder parameters\n");
+        }
+    } else {
+        printf("Received Opus config (informational): bitrate=%u, complexity=%u\n",
+               config.bitrate, config.complexity);
+    }
+
+    return 0;
+}
+
+// IPC MAIN LOOP HELPERS
+
+int audio_common_server_loop(int server_sock, volatile sig_atomic_t *running,
+                             connection_handler_t handler) {
+    while (*running) {
+        printf("Waiting for client connection...\n");
+
+        int client_sock = accept(server_sock, NULL, NULL);
+        if (client_sock < 0) {
+            if (*running) {
+                fprintf(stderr, "Failed to accept client, retrying...\n");
+                struct timespec ts = {1, 0};  // 1 second
+                nanosleep(&ts, NULL);
+                continue;
+            }
+            break;
+        }
+
+        printf("Client connected (fd=%d)\n", client_sock);
+
+        // Run handler with this client
+        handler(client_sock, running);
+
+        // Close client connection
+        close(client_sock);
+
+        if (*running) {
+            printf("Client disconnected, waiting for next client...\n");
+        }
+    }
+
+    return 0;
+}
--- a/internal/audio/c/audio_common.h
+++ b/internal/audio/c/audio_common.h
@ -0,0 +1,160 @@
+/*
+ * JetKVM Audio Common Utilities
+ *
+ * Shared functions used by both audio input and output servers
+ */
+
+#ifndef JETKVM_AUDIO_COMMON_H
+#define JETKVM_AUDIO_COMMON_H
+
+#include <signal.h>
+#include <stdint.h>
+
+// SHARED CONSTANTS
+
+// Audio processing parameters
+#define AUDIO_MAX_PACKET_SIZE     1500     // Maximum Opus packet size
+#define AUDIO_SLEEP_MICROSECONDS  1000     // Default sleep time in microseconds
+#define AUDIO_MAX_ATTEMPTS        5        // Maximum retry attempts
+#define AUDIO_MAX_BACKOFF_US      500000   // Maximum backoff in microseconds
+
+// Error handling
+#define AUDIO_MAX_CONSECUTIVE_ERRORS  10   // Maximum consecutive errors before giving up
+
+// Performance monitoring
+#define AUDIO_TRACE_MASK          0x3FF    // Log every 1024th frame (bit mask for efficiency)
+
+// SIGNAL HANDLERS
+
+/**
+ * Setup signal handlers for graceful shutdown.
+ * Handles SIGTERM and SIGINT by setting the running flag to 0.
+ * Ignores SIGPIPE to prevent crashes on broken pipe writes.
+ *
+ * @param running Pointer to the volatile running flag to set on shutdown
+ */
+void audio_common_setup_signal_handlers(volatile sig_atomic_t *running);
+
+
+/**
+ * Parse integer from environment variable.
+ * Returns default_value if variable is not set or empty.
+ *
+ * @param name Environment variable name
+ * @param default_value Default value if not set
+ * @return Parsed integer value or default
+ */
+int32_t audio_common_parse_env_int(const char *name, int32_t default_value);
+
+/**
+ * Parse string from environment variable.
+ * Returns default_value if variable is not set or empty.
+ *
+ * @param name Environment variable name
+ * @param default_value Default value if not set
+ * @return Environment variable value or default (not duplicated)
+ */
+const char* audio_common_parse_env_string(const char *name, const char *default_value);
+
+
+// COMMON CONFIGURATION
+
+/**
+ * Common audio configuration structure
+ */
+typedef struct {
+    const char *alsa_device;     // ALSA device path
+    int opus_bitrate;            // Opus bitrate
+    int opus_complexity;         // Opus complexity
+    int sample_rate;             // Sample rate
+    int channels;                // Number of channels
+    int frame_size;              // Frame size in samples
+} audio_config_t;
+
+/**
+ * Load common audio configuration from environment
+ * @param config Output configuration
+ * @param is_output true for output server, false for input
+ */
+void audio_common_load_config(audio_config_t *config, int is_output);
+
+/**
+ * Print server startup message
+ * @param server_name Name of the server (e.g., "Audio Output Server")
+ */
+void audio_common_print_startup(const char *server_name);
+
+/**
+ * Print server shutdown message
+ * @param server_name Name of the server
+ */
+void audio_common_print_shutdown(const char *server_name);
+
+// ERROR TRACKING
+
+/**
+ * Error tracking state for audio processing loops
+ */
+typedef struct {
+    uint8_t consecutive_errors;       // Current consecutive error count
+    uint32_t frame_count;             // Total frames processed
+} audio_error_tracker_t;
+
+/**
+ * Initialize error tracker
+ */
+static inline void audio_error_tracker_init(audio_error_tracker_t *tracker) {
+    tracker->consecutive_errors = 0;
+    tracker->frame_count = 0;
+}
+
+/**
+ * Record an error and check if we should give up
+ * Returns 1 if too many errors, 0 to continue
+ */
+static inline uint8_t audio_error_tracker_record_error(audio_error_tracker_t *tracker) {
+    tracker->consecutive_errors++;
+    return (tracker->consecutive_errors >= AUDIO_MAX_CONSECUTIVE_ERRORS) ? 1 : 0;
+}
+
+/**
+ * Record success and increment frame count
+ */
+static inline void audio_error_tracker_record_success(audio_error_tracker_t *tracker) {
+    tracker->consecutive_errors = 0;
+    tracker->frame_count++;
+}
+
+/**
+ * Check if we should log trace info for this frame
+ */
+static inline uint8_t audio_error_tracker_should_trace(audio_error_tracker_t *tracker) {
+    return ((tracker->frame_count & AUDIO_TRACE_MASK) == 1) ? 1 : 0;
+}
+
+
+/**
+ * Parse Opus config message and optionally apply to encoder.
+ * @param data Raw message data
+ * @param length Message length
+ * @param is_encoder If true, apply config to encoder (output server)
+ * @return 0 on success, -1 on error
+ */
+int audio_common_handle_opus_config(const uint8_t *data, uint32_t length, int is_encoder);
+
+// IPC MAIN LOOP HELPERS
+
+/**
+ * Common server accept loop with signal handling.
+ * Accepts clients and calls handler function for each connection.
+ *
+ * @param server_sock Server socket from ipc_create_server
+ * @param running Pointer to running flag (set to 0 on shutdown)
+ * @param handler Connection handler function
+ * @return 0 on clean shutdown, -1 on error
+ */
+typedef int (*connection_handler_t)(int client_sock, volatile sig_atomic_t *running);
+int audio_common_server_loop(int server_sock, volatile sig_atomic_t *running,
+                             connection_handler_t handler);
+
+#endif // JETKVM_AUDIO_COMMON_H
--- a/internal/audio/c/ipc_protocol.c
+++ b/internal/audio/c/ipc_protocol.c
@ -0,0 +1,328 @@
+/*
+ * JetKVM Audio IPC Protocol Implementation
+ *
+ * Implements Unix domain socket communication with exact byte-level
+ * compatibility with Go implementation in internal/audio/ipc_*.go
+ */
+
+#include "ipc_protocol.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <errno.h>
+#include <time.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <sys/uio.h>
+#include <endian.h>
+
+// HELPER FUNCTIONS
+
+/**
+ * Read exactly N bytes from socket (loops until complete or error).
+ * This is critical because read() may return partial data.
+ */
+int ipc_read_full(int sock, void *buf, size_t len) {
+    uint8_t *ptr = (uint8_t *)buf;
+    size_t remaining = len;
+
+    while (remaining > 0) {
+        ssize_t n = read(sock, ptr, remaining);
+
+        if (n < 0) {
+            if (errno == EINTR) {
+                continue;  // Interrupted by signal, retry
+            }
+            return -1;
+        }
+
+        if (n == 0) {
+            return -1;  // Connection closed
+        }
+
+        ptr += n;
+        remaining -= n;
+    }
+
+    return 0;  // Success
+}
+
+
+// MESSAGE READ/WRITE
+
+/**
+ * Read a complete IPC message from socket.
+ * Returns 0 on success, -1 on error.
+ * Caller MUST free msg->data if non-NULL!
+ */
+int ipc_read_message(int sock, ipc_message_t *msg, uint32_t expected_magic) {
+    if (msg == NULL) {
+        return -1;
+    }
+
+    // Initialize message
+    memset(msg, 0, sizeof(ipc_message_t));
+
+    // 1. Read header (9 bytes)
+    if (ipc_read_full(sock, &msg->header, IPC_HEADER_SIZE) != 0) {
+        return -1;
+    }
+
+    // 2. Convert from little-endian (required on big-endian systems)
+    msg->header.magic = le32toh(msg->header.magic);
+    msg->header.length = le32toh(msg->header.length);
+
+    // 3. Validate magic number
+    if (msg->header.magic != expected_magic) {
+        fprintf(stderr, "IPC: Invalid magic number: got 0x%08X, expected 0x%08X\n",
+                msg->header.magic, expected_magic);
+        return -1;
+    }
+
+    // 4. Validate length
+    if (msg->header.length > IPC_MAX_FRAME_SIZE) {
+        fprintf(stderr, "IPC: Message too large: %u bytes (max %d)\n",
+                msg->header.length, IPC_MAX_FRAME_SIZE);
+        return -1;
+    }
+
+    // 5. Read payload if present
+    if (msg->header.length > 0) {
+        msg->data = malloc(msg->header.length);
+        if (msg->data == NULL) {
+            fprintf(stderr, "IPC: Failed to allocate %u bytes for payload\n",
+                    msg->header.length);
+            return -1;
+        }
+
+        if (ipc_read_full(sock, msg->data, msg->header.length) != 0) {
+            free(msg->data);
+            msg->data = NULL;
+            return -1;
+        }
+    }
+
+    return 0;  // Success
+}
+
+/**
+ * Read a complete IPC message using pre-allocated buffer (zero-copy).
+ */
+int ipc_read_message_zerocopy(int sock, ipc_message_t *msg, uint32_t expected_magic,
+                               uint8_t *buffer, uint32_t buffer_size) {
+    if (msg == NULL || buffer == NULL) {
+        return -1;
+    }
+
+    // Initialize message
+    memset(msg, 0, sizeof(ipc_message_t));
+
+    // 1. Read header (9 bytes)
+    if (ipc_read_full(sock, &msg->header, IPC_HEADER_SIZE) != 0) {
+        return -1;
+    }
+
+    // 2. Convert from little-endian
+    msg->header.magic = le32toh(msg->header.magic);
+    msg->header.length = le32toh(msg->header.length);
+
+    // 3. Validate magic number
+    if (msg->header.magic != expected_magic) {
+        return -1;
+    }
+
+    // 4. Validate length
+    if (msg->header.length > IPC_MAX_FRAME_SIZE || msg->header.length > buffer_size) {
+        return -1;
+    }
+
+    // 5. Read payload directly into provided buffer (zero-copy)
+    if (msg->header.length > 0) {
+        if (ipc_read_full(sock, buffer, msg->header.length) != 0) {
+            return -1;
+        }
+        msg->data = buffer;  // Point to provided buffer, no allocation
+    }
+
+    return 0;  // Success
+}
+
+/**
+ * Write a complete IPC message to socket.
+ * Uses writev() for atomic header+payload write.
+ * Returns 0 on success, -1 on error.
+ */
+int ipc_write_message(int sock, uint32_t magic, uint8_t type,
+                      const uint8_t *data, uint32_t length) {
+    // Validate length
+    if (length > IPC_MAX_FRAME_SIZE) {
+        fprintf(stderr, "IPC: Message too large: %u bytes (max %d)\n",
+                length, IPC_MAX_FRAME_SIZE);
+        return -1;
+    }
+
+    // Prepare header
+    ipc_header_t header;
+    header.magic = htole32(magic);
+    header.type = type;
+    header.length = htole32(length);
+
+    // Use writev for atomic write (if possible)
+    struct iovec iov[2];
+    iov[0].iov_base = &header;
+    iov[0].iov_len = IPC_HEADER_SIZE;
+    iov[1].iov_base = (void *)data;
+    iov[1].iov_len = length;
+
+    int iovcnt = (length > 0) ? 2 : 1;
+    size_t total_len = IPC_HEADER_SIZE + length;
+
+    ssize_t written = writev(sock, iov, iovcnt);
+
+    if (written < 0) {
+        if (errno == EINTR) {
+            // Retry once on interrupt
+            written = writev(sock, iov, iovcnt);
+        }
+
+        if (written < 0) {
+            perror("IPC: writev failed");
+            return -1;
+        }
+    }
+
+    if ((size_t)written != total_len) {
+        fprintf(stderr, "IPC: Partial write: %zd/%zu bytes\n", written, total_len);
+        return -1;
+    }
+
+    return 0;  // Success
+}
+
+
+/**
+ * Parse Opus configuration from message data (36 bytes, little-endian).
+ */
+int ipc_parse_opus_config(const uint8_t *data, uint32_t length, ipc_opus_config_t *config) {
+    if (data == NULL || config == NULL) {
+        return -1;
+    }
+
+    if (length != 36) {
+        fprintf(stderr, "IPC: Invalid Opus config size: %u bytes (expected 36)\n", length);
+        return -1;
+    }
+
+    // Parse little-endian uint32 fields
+    const uint32_t *u32_data = (const uint32_t *)data;
+    config->sample_rate = le32toh(u32_data[0]);
+    config->channels = le32toh(u32_data[1]);
+    config->frame_size = le32toh(u32_data[2]);
+    config->bitrate = le32toh(u32_data[3]);
+    config->complexity = le32toh(u32_data[4]);
+    config->vbr = le32toh(u32_data[5]);
+    config->signal_type = le32toh(u32_data[6]);
+    config->bandwidth = le32toh(u32_data[7]);
+    config->dtx = le32toh(u32_data[8]);
+
+    return 0;  // Success
+}
+
+/**
+ * Parse basic audio configuration from message data (12 bytes, little-endian).
+ */
+int ipc_parse_config(const uint8_t *data, uint32_t length, ipc_config_t *config) {
+    if (data == NULL || config == NULL) {
+        return -1;
+    }
+
+    if (length != 12) {
+        fprintf(stderr, "IPC: Invalid config size: %u bytes (expected 12)\n", length);
+        return -1;
+    }
+
+    // Parse little-endian uint32 fields
+    const uint32_t *u32_data = (const uint32_t *)data;
+    config->sample_rate = le32toh(u32_data[0]);
+    config->channels = le32toh(u32_data[1]);
+    config->frame_size = le32toh(u32_data[2]);
+
+    return 0;  // Success
+}
+
+/**
+ * Free message resources.
+ */
+void ipc_free_message(ipc_message_t *msg) {
+    if (msg != NULL && msg->data != NULL) {
+        free(msg->data);
+        msg->data = NULL;
+    }
+}
+
+// SOCKET MANAGEMENT
+
+/**
+ * Create Unix domain socket server.
+ */
+int ipc_create_server(const char *socket_path) {
+    if (socket_path == NULL) {
+        return -1;
+    }
+
+    // 1. Create socket
+    int sock = socket(AF_UNIX, SOCK_STREAM, 0);
+    if (sock < 0) {
+        perror("IPC: socket() failed");
+        return -1;
+    }
+
+    // 2. Remove existing socket file (ignore errors)
+    unlink(socket_path);
+
+    // 3. Bind to path
+    struct sockaddr_un addr;
+    memset(&addr, 0, sizeof(addr));
+    addr.sun_family = AF_UNIX;
+
+    if (strlen(socket_path) >= sizeof(addr.sun_path)) {
+        fprintf(stderr, "IPC: Socket path too long: %s\n", socket_path);
+        close(sock);
+        return -1;
+    }
+
+    strncpy(addr.sun_path, socket_path, sizeof(addr.sun_path) - 1);
+
+    if (bind(sock, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
+        perror("IPC: bind() failed");
+        close(sock);
+        return -1;
+    }
+
+    // 4. Listen with backlog=1 (single client)
+    if (listen(sock, 1) < 0) {
+        perror("IPC: listen() failed");
+        close(sock);
+        return -1;
+    }
+
+    printf("IPC: Server listening on %s\n", socket_path);
+    return sock;
+}
+
+/**
+ * Accept client connection.
+ */
+int ipc_accept_client(int server_sock) {
+    int client_sock = accept(server_sock, NULL, NULL);
+
+    if (client_sock < 0) {
+        perror("IPC: accept() failed");
+        return -1;
+    }
+
+    printf("IPC: Client connected (fd=%d)\n", client_sock);
+    return client_sock;
+}
--- a/internal/audio/c/ipc_protocol.h
+++ b/internal/audio/c/ipc_protocol.h
@ -0,0 +1,211 @@
+/*
+ * JetKVM Audio IPC Protocol
+ *
+ * Wire protocol for Unix domain socket communication between main process
+ * and audio subprocesses. This protocol is 100% compatible with the Go
+ * implementation in internal/audio/ipc_*.go
+ *
+ * CRITICAL: All multi-byte integers use LITTLE-ENDIAN byte order.
+ */
+
+#ifndef JETKVM_IPC_PROTOCOL_H
+#define JETKVM_IPC_PROTOCOL_H
+
+#include <stdint.h>
+#include <sys/types.h>
+
+// PROTOCOL CONSTANTS
+
+// Magic numbers (ASCII representation when read as little-endian)
+#define IPC_MAGIC_OUTPUT 0x4A4B4F55  // "JKOU" - JetKVM Output (device → browser)
+#define IPC_MAGIC_INPUT  0x4A4B4D49  // "JKMI" - JetKVM Microphone Input (browser → device)
+
+// Message types (matches Go UnifiedMessageType enum)
+#define IPC_MSG_TYPE_OPUS_FRAME  0  // Audio frame data (Opus encoded)
+#define IPC_MSG_TYPE_CONFIG      1  // Basic audio config (12 bytes)
+#define IPC_MSG_TYPE_OPUS_CONFIG 2  // Complete Opus config (36 bytes)
+#define IPC_MSG_TYPE_STOP        3  // Shutdown signal
+#define IPC_MSG_TYPE_HEARTBEAT   4  // Keep-alive ping
+#define IPC_MSG_TYPE_ACK         5  // Acknowledgment
+
+// Size constraints
+#define IPC_HEADER_SIZE 9         // Fixed header size (reduced from 17)
+#define IPC_MAX_FRAME_SIZE 1024   // Maximum payload size (128kbps @ 20ms = ~600 bytes worst case with VBR+FEC)
+
+// Socket paths
+#define IPC_SOCKET_OUTPUT "/var/run/audio_output.sock"
+#define IPC_SOCKET_INPUT  "/var/run/audio_input.sock"
+
+// WIRE FORMAT STRUCTURES
+
+/**
+ * IPC message header (9 bytes, little-endian)
+ *
+ * Byte layout:
+ * [0-3]   magic      uint32_t LE  Magic number (0x4A4B4F55 or 0x4A4B4D49)
+ * [4]     type       uint8_t      Message type (0-5)
+ * [5-8]   length     uint32_t LE  Payload size in bytes
+ * [9+]    data       uint8_t[]    Variable payload
+ *
+ * CRITICAL: Must use __attribute__((packed)) to prevent padding.
+ *
+ * NOTE: Timestamp removed (was unused, saved 8 bytes per message)
+ */
+typedef struct __attribute__((packed)) {
+    uint32_t magic;      // Magic number (LE)
+    uint8_t  type;       // Message type
+    uint32_t length;     // Payload length in bytes (LE)
+} ipc_header_t;
+
+/**
+ * Basic audio configuration (12 bytes)
+ * Message type: IPC_MSG_TYPE_CONFIG
+ *
+ * All fields are uint32_t little-endian.
+ */
+typedef struct __attribute__((packed)) {
+    uint32_t sample_rate;  // Samples per second (e.g., 48000)
+    uint32_t channels;     // Number of channels (e.g., 2 for stereo)
+    uint32_t frame_size;   // Samples per frame (e.g., 960)
+} ipc_config_t;
+
+/**
+ * Complete Opus encoder/decoder configuration (36 bytes)
+ * Message type: IPC_MSG_TYPE_OPUS_CONFIG
+ *
+ * All fields are uint32_t little-endian.
+ * Note: Negative values (like signal_type=-1000) are stored as two's complement uint32.
+ */
+typedef struct __attribute__((packed)) {
+    uint32_t sample_rate;  // Samples per second (48000)
+    uint32_t channels;     // Number of channels (2)
+    uint32_t frame_size;   // Samples per frame per channel (960 = 20ms @ 48kHz)
+    uint32_t bitrate;      // Bits per second (128000)
+    uint32_t complexity;   // Encoder complexity 0-10 (2=balanced quality/speed)
+    uint32_t vbr;          // Variable bitrate: 0=disabled, 1=enabled
+    uint32_t signal_type;  // Signal type: -1000=auto, 3001=voice, 3002=music
+    uint32_t bandwidth;    // Bandwidth: 1101=narrowband, 1102=mediumband, 1103=wideband, 1104=superwideband, 1105=fullband
+    uint32_t dtx;          // Discontinuous transmission: 0=disabled, 1=enabled
+} ipc_opus_config_t;
+
+/**
+ * Complete IPC message (header + payload)
+ */
+typedef struct {
+    ipc_header_t header;
+    uint8_t *data;         // Dynamically allocated payload (NULL if length=0)
+} ipc_message_t;
+
+
+/**
+ * Read a complete IPC message from socket.
+ *
+ * This function:
+ * 1. Reads exactly 9 bytes (header)
+ * 2. Validates magic number
+ * 3. Validates length <= IPC_MAX_FRAME_SIZE
+ * 4. Allocates and reads payload if length > 0
+ * 5. Stores result in msg->header and msg->data
+ *
+ * @param sock Socket file descriptor
+ * @param msg Output message (data will be malloc'd if length > 0)
+ * @param expected_magic Expected magic number (IPC_MAGIC_OUTPUT or IPC_MAGIC_INPUT)
+ * @return 0 on success, -1 on error
+ *
+ * CALLER MUST FREE msg->data if non-NULL!
+ */
+int ipc_read_message(int sock, ipc_message_t *msg, uint32_t expected_magic);
+
+/**
+ * Read a complete IPC message using pre-allocated buffer (zero-copy).
+ *
+ * @param sock Socket file descriptor
+ * @param msg Message structure to fill
+ * @param expected_magic Expected magic number for validation
+ * @param buffer Pre-allocated buffer for message data
+ * @param buffer_size Size of pre-allocated buffer
+ * @return 0 on success, -1 on error
+ *
+ * msg->data will point to buffer (no allocation). Caller does NOT need to free.
+ */
+int ipc_read_message_zerocopy(int sock, ipc_message_t *msg, uint32_t expected_magic,
+                               uint8_t *buffer, uint32_t buffer_size);
+
+/**
+ * Write a complete IPC message to socket.
+ *
+ * This function writes header + payload atomically (if possible via writev).
+ *
+ * @param sock Socket file descriptor
+ * @param magic Magic number (IPC_MAGIC_OUTPUT or IPC_MAGIC_INPUT)
+ * @param type Message type (IPC_MSG_TYPE_*)
+ * @param data Payload data (can be NULL if length=0)
+ * @param length Payload length in bytes
+ * @return 0 on success, -1 on error
+ */
+int ipc_write_message(int sock, uint32_t magic, uint8_t type,
+                      const uint8_t *data, uint32_t length);
+
+/**
+ * Parse Opus configuration from message data.
+ *
+ * @param data Payload data (must be exactly 36 bytes)
+ * @param length Payload length (must be 36)
+ * @param config Output Opus configuration
+ * @return 0 on success, -1 if length != 36
+ */
+int ipc_parse_opus_config(const uint8_t *data, uint32_t length, ipc_opus_config_t *config);
+
+/**
+ * Parse basic audio configuration from message data.
+ *
+ * @param data Payload data (must be exactly 12 bytes)
+ * @param length Payload length (must be 12)
+ * @param config Output audio configuration
+ * @return 0 on success, -1 if length != 12
+ */
+int ipc_parse_config(const uint8_t *data, uint32_t length, ipc_config_t *config);
+
+/**
+ * Free message resources.
+ *
+ * @param msg Message to free (frees msg->data if non-NULL)
+ */
+void ipc_free_message(ipc_message_t *msg);
+
+
+/**
+ * Create Unix domain socket server.
+ *
+ * This function:
+ * 1. Creates socket with AF_UNIX, SOCK_STREAM
+ * 2. Removes existing socket file
+ * 3. Binds to specified path
+ * 4. Listens with backlog=1 (single client)
+ *
+ * @param socket_path Path to Unix socket (e.g., "/var/run/audio_output.sock")
+ * @return Socket fd on success, -1 on error
+ */
+int ipc_create_server(const char *socket_path);
+
+/**
+ * Accept client connection with automatic retry.
+ *
+ * Blocks until client connects or error occurs.
+ *
+ * @param server_sock Server socket fd from ipc_create_server()
+ * @return Client socket fd on success, -1 on error
+ */
+int ipc_accept_client(int server_sock);
+
+/**
+ * Helper: Read exactly N bytes from socket (loops until complete or error).
+ *
+ * @param sock Socket file descriptor
+ * @param buf Output buffer
+ * @param len Number of bytes to read
+ * @return 0 on success, -1 on error
+ */
+int ipc_read_full(int sock, void *buf, size_t len);
+
+#endif // JETKVM_IPC_PROTOCOL_H
--- a/internal/audio/c/jetkvm_audio_input.c
+++ b/internal/audio/c/jetkvm_audio_input.c
@ -0,0 +1,169 @@
+/*
+ * JetKVM Audio Input Server
+ *
+ * Standalone C binary for audio input path:
+ * Browser → WebRTC → Go Process → IPC Receive → Opus Decode → ALSA Playback (USB Gadget)
+ *
+ * This replaces the Go subprocess that was running with --audio-input-server flag.
+ *
+ * IMPORTANT: This binary only does OPUS DECODING (not encoding).
+ * The browser already encodes audio to Opus before sending via WebRTC.
+ */
+
+#include "ipc_protocol.h"
+#include "audio_common.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <signal.h>
+#include <errno.h>
+
+// Forward declarations from audio.c
+extern int jetkvm_audio_playback_init(void);
+extern void jetkvm_audio_playback_close(void);
+extern int jetkvm_audio_decode_write(void *opus_buf, int opus_size);
+extern void update_audio_decoder_constants(uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                                           uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff);
+
+
+static volatile sig_atomic_t g_running = 1;
+
+
+/**
+ * Send ACK response for heartbeat messages.
+ */
+static inline int32_t send_ack(int32_t client_sock) {
+    return ipc_write_message(client_sock, IPC_MAGIC_INPUT, IPC_MSG_TYPE_ACK, NULL, 0);
+}
+
+
+/**
+ * Main audio decode and playback loop.
+ * Receives Opus frames via IPC, decodes, writes to ALSA.
+ */
+static int run_audio_loop(int client_sock, volatile sig_atomic_t *running) {
+    audio_error_tracker_t tracker;
+    audio_error_tracker_init(&tracker);
+
+    // Static buffer for zero-copy IPC (no malloc/free per frame)
+    static uint8_t frame_buffer[IPC_MAX_FRAME_SIZE] __attribute__((aligned(64)));
+
+    printf("Starting audio input loop...\n");
+
+    while (*running) {
+        ipc_message_t msg;
+
+        if (ipc_read_message_zerocopy(client_sock, &msg, IPC_MAGIC_INPUT,
+                                      frame_buffer, sizeof(frame_buffer)) != 0) {
+            if (*running) {
+                fprintf(stderr, "Failed to read message from client\n");
+            }
+            break;
+        }
+
+        switch (msg.header.type) {
+            case IPC_MSG_TYPE_OPUS_FRAME: {
+                if (msg.header.length == 0 || msg.data == NULL) {
+                    fprintf(stderr, "Warning: Empty Opus frame received\n");
+                    continue;
+                }
+
+                int frames_written = jetkvm_audio_decode_write(msg.data, msg.header.length);
+
+                if (frames_written < 0) {
+                    fprintf(stderr, "Audio decode/write failed (error %d/%d)\n",
+                            tracker.consecutive_errors + 1, AUDIO_MAX_CONSECUTIVE_ERRORS);
+
+                    if (audio_error_tracker_record_error(&tracker)) {
+                        fprintf(stderr, "Too many consecutive errors, giving up\n");
+                        return -1;
+                    }
+                } else {
+                    audio_error_tracker_record_success(&tracker);
+
+                    if (audio_error_tracker_should_trace(&tracker)) {
+                        printf("Processed frame %u (opus_size=%u, pcm_frames=%d)\n",
+                               tracker.frame_count, msg.header.length, frames_written);
+                    }
+                }
+
+                break;
+            }
+
+            case IPC_MSG_TYPE_CONFIG:
+                printf("Received basic audio config\n");
+                send_ack(client_sock);
+                break;
+
+            case IPC_MSG_TYPE_OPUS_CONFIG:
+                audio_common_handle_opus_config(msg.data, msg.header.length, 0);
+                send_ack(client_sock);
+                break;
+
+            case IPC_MSG_TYPE_STOP:
+                printf("Received stop message\n");
+                *running = 0;
+                return 0;
+
+            case IPC_MSG_TYPE_HEARTBEAT:
+                send_ack(client_sock);
+                break;
+
+            default:
+                printf("Warning: Unknown message type: %u\n", msg.header.type);
+                break;
+        }
+    }
+
+    printf("Audio input loop ended after %u frames\n", tracker.frame_count);
+    return 0;
+}
+
+
+int main(int argc, char **argv) {
+    audio_common_print_startup("Audio Input Server");
+
+    // Setup signal handlers
+    audio_common_setup_signal_handlers(&g_running);
+
+    // Load configuration from environment
+    audio_config_t config;
+    audio_common_load_config(&config, 0);  // 0 = input server
+
+    // Apply decoder constants to audio.c (encoder params not needed)
+    update_audio_decoder_constants(
+        config.sample_rate,
+        config.channels,
+        config.frame_size,
+        AUDIO_MAX_PACKET_SIZE,
+        AUDIO_SLEEP_MICROSECONDS,
+        AUDIO_MAX_ATTEMPTS,
+        AUDIO_MAX_BACKOFF_US
+    );
+
+    // Initialize audio playback (Opus decoder + ALSA playback)
+    printf("Initializing audio playback on device: %s\n", config.alsa_device);
+    if (jetkvm_audio_playback_init() != 0) {
+        fprintf(stderr, "Failed to initialize audio playback\n");
+        return 1;
+    }
+
+    // Create IPC server
+    int server_sock = ipc_create_server(IPC_SOCKET_INPUT);
+    if (server_sock < 0) {
+        fprintf(stderr, "Failed to create IPC server\n");
+        jetkvm_audio_playback_close();
+        return 1;
+    }
+
+    // Main connection loop
+    audio_common_server_loop(server_sock, &g_running, run_audio_loop);
+
+    audio_common_print_shutdown("audio input server");
+    close(server_sock);
+    unlink(IPC_SOCKET_INPUT);
+    jetkvm_audio_playback_close();
+
+    printf("Audio input server exited cleanly\n");
+    return 0;
+}
--- a/internal/audio/c/jetkvm_audio_output.c
+++ b/internal/audio/c/jetkvm_audio_output.c
@ -0,0 +1,193 @@
+/*
+ * JetKVM Audio Output Server
+ *
+ * Standalone C binary for audio output path:
+ * ALSA Capture (TC358743 HDMI) → Opus Encode → IPC Send → Go Process → WebRTC → Browser
+ *
+ * This replaces the Go subprocess that was running with --audio-output-server flag.
+ */
+
+#include "ipc_protocol.h"
+#include "audio_common.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <signal.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <sched.h>
+#include <time.h>
+
+// Forward declarations from audio.c
+extern int jetkvm_audio_capture_init(void);
+extern void jetkvm_audio_capture_close(void);
+extern int jetkvm_audio_read_encode(void *opus_buf);
+extern void update_audio_constants(uint32_t bitrate, uint8_t complexity,
+                                   uint32_t sr, uint8_t ch, uint16_t fs, uint16_t max_pkt,
+                                   uint32_t sleep_us, uint8_t max_attempts, uint32_t max_backoff);
+extern int update_opus_encoder_params(uint32_t bitrate, uint8_t complexity);
+
+
+static volatile sig_atomic_t g_running = 1;
+
+
+static void load_output_config(audio_config_t *common) {
+    audio_common_load_config(common, 1);  // 1 = output server
+}
+
+
+/**
+ * Handle incoming IPC messages from client (non-blocking).
+ * Returns 0 on success, -1 on error.
+ */
+static int handle_incoming_messages(int client_sock, volatile sig_atomic_t *running) {
+    // Static buffer for zero-copy IPC (control messages are small)
+    static uint8_t msg_buffer[IPC_MAX_FRAME_SIZE] __attribute__((aligned(64)));
+
+    // Set non-blocking mode for client socket
+    int flags = fcntl(client_sock, F_GETFL, 0);
+    fcntl(client_sock, F_SETFL, flags | O_NONBLOCK);
+
+    ipc_message_t msg;
+
+    // Try to read message (non-blocking, zero-copy)
+    int result = ipc_read_message_zerocopy(client_sock, &msg, IPC_MAGIC_OUTPUT,
+                                           msg_buffer, sizeof(msg_buffer));
+
+    // Restore blocking mode
+    fcntl(client_sock, F_SETFL, flags);
+
+    if (result != 0) {
+        if (errno == EAGAIN || errno == EWOULDBLOCK) {
+            return 0;  // No message available, not an error
+        }
+        return -1;
+    }
+
+    switch (msg.header.type) {
+        case IPC_MSG_TYPE_OPUS_CONFIG:
+            audio_common_handle_opus_config(msg.data, msg.header.length, 1);
+            break;
+
+        case IPC_MSG_TYPE_STOP:
+            printf("Received stop message\n");
+            *running = 0;
+            break;
+
+        case IPC_MSG_TYPE_HEARTBEAT:
+            break;
+
+        default:
+            printf("Warning: Unknown message type: %u\n", msg.header.type);
+            break;
+    }
+
+    return 0;
+}
+
+
+/**
+ * Main audio capture and encode loop.
+ * Continuously reads from ALSA, encodes to Opus, sends via IPC.
+ */
+static int run_audio_loop(int client_sock, volatile sig_atomic_t *running) {
+    uint8_t opus_buffer[IPC_MAX_FRAME_SIZE];
+    audio_error_tracker_t tracker;
+    audio_error_tracker_init(&tracker);
+
+    printf("Starting audio output loop...\n");
+
+    while (*running) {
+        if (handle_incoming_messages(client_sock, running) < 0) {
+            fprintf(stderr, "Client disconnected, waiting for reconnection...\n");
+            break;
+        }
+
+        int opus_size = jetkvm_audio_read_encode(opus_buffer);
+
+        if (opus_size < 0) {
+            fprintf(stderr, "Audio read/encode failed (error %d/%d)\n",
+                    tracker.consecutive_errors + 1, AUDIO_MAX_CONSECUTIVE_ERRORS);
+
+            if (audio_error_tracker_record_error(&tracker)) {
+                fprintf(stderr, "Too many consecutive errors, giving up\n");
+                return -1;
+            }
+            // No sleep needed - jetkvm_audio_read_encode already uses snd_pcm_wait internally
+            continue;
+        }
+
+        if (opus_size == 0) {
+            // Frame skipped for recovery, minimal yield
+            sched_yield();
+            continue;
+        }
+
+        audio_error_tracker_record_success(&tracker);
+
+        if (ipc_write_message(client_sock, IPC_MAGIC_OUTPUT, IPC_MSG_TYPE_OPUS_FRAME,
+                             opus_buffer, opus_size) != 0) {
+            fprintf(stderr, "Failed to send frame to client\n");
+            break;
+        }
+
+        if (audio_error_tracker_should_trace(&tracker)) {
+            printf("Sent frame %u (size=%d bytes)\n", tracker.frame_count, opus_size);
+        }
+    }
+
+    printf("Audio output loop ended after %u frames\n", tracker.frame_count);
+    return 0;
+}
+
+
+int main(int argc, char **argv) {
+    audio_common_print_startup("Audio Output Server");
+
+    // Setup signal handlers
+    audio_common_setup_signal_handlers(&g_running);
+
+    // Load configuration from environment
+    audio_config_t common;
+    load_output_config(&common);
+
+    // Apply audio constants to audio.c
+    update_audio_constants(
+        common.opus_bitrate,
+        common.opus_complexity,
+        common.sample_rate,
+        common.channels,
+        common.frame_size,
+        AUDIO_MAX_PACKET_SIZE,
+        AUDIO_SLEEP_MICROSECONDS,
+        AUDIO_MAX_ATTEMPTS,
+        AUDIO_MAX_BACKOFF_US
+    );
+
+    // Initialize audio capture
+    printf("Initializing audio capture on device: %s\n", common.alsa_device);
+    if (jetkvm_audio_capture_init() != 0) {
+        fprintf(stderr, "Failed to initialize audio capture\n");
+        return 1;
+    }
+
+    // Create IPC server
+    int server_sock = ipc_create_server(IPC_SOCKET_OUTPUT);
+    if (server_sock < 0) {
+        fprintf(stderr, "Failed to create IPC server\n");
+        jetkvm_audio_capture_close();
+        return 1;
+    }
+
+    // Main connection loop
+    audio_common_server_loop(server_sock, &g_running, run_audio_loop);
+
+    audio_common_print_shutdown("audio output server");
+    close(server_sock);
+    unlink(IPC_SOCKET_OUTPUT);
+    jetkvm_audio_capture_close();
+
+    printf("Audio output server exited cleanly\n");
+    return 0;
+}
--- a/internal/audio/embed.go
+++ b/internal/audio/embed.go
@ -0,0 +1,80 @@
+package audio
+
+import (
+	_ "embed"
+	"fmt"
+	"os"
+)
+
+// Embedded C audio binaries (built during compilation)
+//
+//go:embed bin/jetkvm_audio_output
+var audioOutputBinary []byte
+
+//go:embed bin/jetkvm_audio_input
+var audioInputBinary []byte
+
+const (
+	audioBinDir        = "/userdata/jetkvm/bin"
+	audioOutputBinPath = audioBinDir + "/jetkvm_audio_output"
+	audioInputBinPath  = audioBinDir + "/jetkvm_audio_input"
+	binaryFileMode     = 0755 // rwxr-xr-x
+)
+
+// ExtractEmbeddedBinaries extracts the embedded C audio binaries to disk
+// This should be called during application startup before audio supervisors are started
+func ExtractEmbeddedBinaries() error {
+	// Create bin directory if it doesn't exist
+	if err := os.MkdirAll(audioBinDir, 0755); err != nil {
+		return fmt.Errorf("failed to create audio bin directory: %w", err)
+	}
+
+	// Extract audio output binary
+	if err := extractBinary(audioOutputBinary, audioOutputBinPath); err != nil {
+		return fmt.Errorf("failed to extract audio output binary: %w", err)
+	}
+
+	// Extract audio input binary
+	if err := extractBinary(audioInputBinary, audioInputBinPath); err != nil {
+		return fmt.Errorf("failed to extract audio input binary: %w", err)
+	}
+
+	return nil
+}
+
+// extractBinary writes embedded binary data to disk with executable permissions
+func extractBinary(data []byte, path string) error {
+	// Check if binary already exists and is valid
+	if info, err := os.Stat(path); err == nil {
+		// File exists - check if size matches
+		if info.Size() == int64(len(data)) {
+			// Binary already extracted and matches embedded version
+			return nil
+		}
+		// Size mismatch - need to update
+	}
+
+	// Write to temporary file first for atomic replacement
+	tmpPath := path + ".tmp"
+	if err := os.WriteFile(tmpPath, data, binaryFileMode); err != nil {
+		return fmt.Errorf("failed to write binary to %s: %w", tmpPath, err)
+	}
+
+	// Atomically rename to final path
+	if err := os.Rename(tmpPath, path); err != nil {
+		os.Remove(tmpPath) // Clean up on error
+		return fmt.Errorf("failed to rename binary to %s: %w", path, err)
+	}
+
+	return nil
+}
+
+// GetAudioOutputBinaryPath returns the path to the audio output binary
+func GetAudioOutputBinaryPath() string {
+	return audioOutputBinPath
+}
+
+// GetAudioInputBinaryPath returns the path to the audio input binary
+func GetAudioInputBinaryPath() string {
+	return audioInputBinPath
+}
--- a/internal/audio/ipc.go
+++ b/internal/audio/ipc.go
@ -0,0 +1,185 @@
+package audio
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+	"net"
+	"sync"
+	"time"
+
+	"github.com/jetkvm/kvm/internal/logging"
+	"github.com/rs/zerolog"
+)
+
+// Buffer pool for zero-allocation writes
+var writeBufferPool = sync.Pool{
+	New: func() interface{} {
+		buf := make([]byte, ipcHeaderSize+ipcMaxFrameSize)
+		return &buf
+	},
+}
+
+// IPC Protocol constants (matches C implementation in ipc_protocol.h)
+const (
+	ipcMagicOutput   = 0x4A4B4F55 // "JKOU" - Output (device → browser)
+	ipcMagicInput    = 0x4A4B4D49 // "JKMI" - Input (browser → device)
+	ipcHeaderSize    = 9          // Reduced from 17 (removed 8-byte timestamp)
+	ipcMaxFrameSize  = 1024       // 128kbps @ 20ms = ~600 bytes worst case with VBR+FEC
+	ipcMsgTypeOpus   = 0
+	ipcMsgTypeConfig = 1
+	ipcMsgTypeStop   = 3
+	connectTimeout   = 5 * time.Second
+	readTimeout      = 2 * time.Second
+)
+
+// IPCClient manages Unix socket communication with audio subprocess
+type IPCClient struct {
+	socketPath  string
+	magicNumber uint32
+	conn        net.Conn
+	mu          sync.Mutex
+	logger      zerolog.Logger
+	readBuf     []byte // Reusable buffer for reads (single reader per client)
+}
+
+// NewIPCClient creates a new IPC client
+// For output: socketPath="/var/run/audio_output.sock", magic=ipcMagicOutput
+// For input:  socketPath="/var/run/audio_input.sock", magic=ipcMagicInput
+func NewIPCClient(name, socketPath string, magicNumber uint32) *IPCClient {
+	logger := logging.GetDefaultLogger().With().Str("component", name+"-ipc").Logger()
+
+	return &IPCClient{
+		socketPath:  socketPath,
+		magicNumber: magicNumber,
+		logger:      logger,
+		readBuf:     make([]byte, ipcMaxFrameSize),
+	}
+}
+
+// Connect establishes connection to the subprocess
+func (c *IPCClient) Connect() error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.conn != nil {
+		c.conn.Close()
+		c.conn = nil
+	}
+
+	conn, err := net.DialTimeout("unix", c.socketPath, connectTimeout)
+	if err != nil {
+		return fmt.Errorf("failed to connect to %s: %w", c.socketPath, err)
+	}
+
+	c.conn = conn
+	c.logger.Debug().Str("socket", c.socketPath).Msg("connected to subprocess")
+	return nil
+}
+
+// Disconnect closes the connection
+func (c *IPCClient) Disconnect() {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.conn != nil {
+		c.conn.Close()
+		c.conn = nil
+		c.logger.Debug().Msg("disconnected from subprocess")
+	}
+}
+
+// IsConnected returns true if currently connected
+func (c *IPCClient) IsConnected() bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.conn != nil
+}
+
+// ReadMessage reads a complete IPC message (header + payload)
+// Returns message type, payload data, and error
+// IMPORTANT: The returned payload slice is only valid until the next ReadMessage call.
+// Callers must use the data immediately or copy if retention is needed.
+func (c *IPCClient) ReadMessage() (uint8, []byte, error) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.conn == nil {
+		return 0, nil, fmt.Errorf("not connected")
+	}
+
+	// Set read deadline
+	if err := c.conn.SetReadDeadline(time.Now().Add(readTimeout)); err != nil {
+		return 0, nil, fmt.Errorf("failed to set read deadline: %w", err)
+	}
+
+	// Read 9-byte header
+	var header [ipcHeaderSize]byte
+	if _, err := io.ReadFull(c.conn, header[:]); err != nil {
+		return 0, nil, fmt.Errorf("failed to read header: %w", err)
+	}
+
+	// Parse header (little-endian)
+	magic := binary.LittleEndian.Uint32(header[0:4])
+	msgType := header[4]
+	length := binary.LittleEndian.Uint32(header[5:9])
+
+	// Validate magic number
+	if magic != c.magicNumber {
+		return 0, nil, fmt.Errorf("invalid magic: got 0x%X, expected 0x%X", magic, c.magicNumber)
+	}
+
+	// Validate length
+	if length > ipcMaxFrameSize {
+		return 0, nil, fmt.Errorf("message too large: %d bytes", length)
+	}
+
+	// Read payload if present
+	if length == 0 {
+		return msgType, nil, nil
+	}
+
+	// Read directly into reusable buffer (zero-allocation)
+	if _, err := io.ReadFull(c.conn, c.readBuf[:length]); err != nil {
+		return 0, nil, fmt.Errorf("failed to read payload: %w", err)
+	}
+
+	// Return slice of readBuf - caller must use immediately, data is only valid until next ReadMessage
+	// This avoids allocation in hot path (50 frames/sec)
+	return msgType, c.readBuf[:length], nil
+}
+
+// WriteMessage writes a complete IPC message
+func (c *IPCClient) WriteMessage(msgType uint8, payload []byte) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.conn == nil {
+		return fmt.Errorf("not connected")
+	}
+
+	length := uint32(len(payload))
+	if length > ipcMaxFrameSize {
+		return fmt.Errorf("payload too large: %d bytes", length)
+	}
+
+	// Get buffer from pool for zero-allocation write
+	bufPtr := writeBufferPool.Get().(*[]byte)
+	defer writeBufferPool.Put(bufPtr)
+	buf := *bufPtr
+
+	// Build header in pooled buffer (9 bytes, little-endian)
+	binary.LittleEndian.PutUint32(buf[0:4], c.magicNumber)
+	buf[4] = msgType
+	binary.LittleEndian.PutUint32(buf[5:9], length)
+
+	// Copy payload after header
+	copy(buf[ipcHeaderSize:], payload)
+
+	// Write header + payload atomically
+	if _, err := c.conn.Write(buf[:ipcHeaderSize+length]); err != nil {
+		return fmt.Errorf("failed to write message: %w", err)
+	}
+
+	return nil
+}
--- a/internal/audio/relay.go
+++ b/internal/audio/relay.go
@ -0,0 +1,152 @@
+package audio
+
+import (
+	"context"
+	"fmt"
+	"sync/atomic"
+	"time"
+
+	"github.com/jetkvm/kvm/internal/logging"
+	"github.com/pion/webrtc/v4"
+	"github.com/pion/webrtc/v4/pkg/media"
+	"github.com/rs/zerolog"
+)
+
+// OutputRelay forwards audio from subprocess (HDMI) to WebRTC (browser)
+type OutputRelay struct {
+	client     *IPCClient
+	audioTrack *webrtc.TrackLocalStaticSample
+	ctx        context.Context
+	cancel     context.CancelFunc
+	logger     zerolog.Logger
+	running    atomic.Bool
+	sample     media.Sample // Reusable sample for zero-allocation hot path
+
+	// Stats (Uint32: overflows after 2.7 years @ 50fps, faster atomics on 32-bit ARM)
+	framesRelayed atomic.Uint32
+	framesDropped atomic.Uint32
+}
+
+// NewOutputRelay creates a relay for output audio (device → browser)
+func NewOutputRelay(client *IPCClient, audioTrack *webrtc.TrackLocalStaticSample) *OutputRelay {
+	ctx, cancel := context.WithCancel(context.Background())
+	logger := logging.GetDefaultLogger().With().Str("component", "audio-output-relay").Logger()
+
+	return &OutputRelay{
+		client:     client,
+		audioTrack: audioTrack,
+		ctx:        ctx,
+		cancel:     cancel,
+		logger:     logger,
+		sample: media.Sample{
+			Duration: 20 * time.Millisecond, // Constant for all Opus frames
+		},
+	}
+}
+
+// Start begins relaying audio frames
+func (r *OutputRelay) Start() error {
+	if r.running.Swap(true) {
+		return fmt.Errorf("output relay already running")
+	}
+
+	go r.relayLoop()
+	r.logger.Debug().Msg("output relay started")
+	return nil
+}
+
+// Stop stops the relay
+func (r *OutputRelay) Stop() {
+	if !r.running.Swap(false) {
+		return
+	}
+
+	r.cancel()
+	r.logger.Debug().
+		Uint32("frames_relayed", r.framesRelayed.Load()).
+		Uint32("frames_dropped", r.framesDropped.Load()).
+		Msg("output relay stopped")
+}
+
+// relayLoop continuously reads from IPC and writes to WebRTC
+func (r *OutputRelay) relayLoop() {
+	const reconnectDelay = 1 * time.Second
+
+	for r.running.Load() {
+		// Ensure connected
+		if !r.client.IsConnected() {
+			if err := r.client.Connect(); err != nil {
+				r.logger.Debug().Err(err).Msg("failed to connect, will retry")
+				time.Sleep(reconnectDelay)
+				continue
+			}
+		}
+
+		// Read message from subprocess
+		msgType, payload, err := r.client.ReadMessage()
+		if err != nil {
+			// Connection error - reconnect
+			if r.running.Load() {
+				r.logger.Warn().Err(err).Msg("read error, reconnecting")
+				r.client.Disconnect()
+				time.Sleep(reconnectDelay)
+			}
+			continue
+		}
+
+		// Handle message
+		if msgType == ipcMsgTypeOpus && len(payload) > 0 {
+			// Reuse sample struct (zero-allocation hot path)
+			r.sample.Data = payload
+
+			if err := r.audioTrack.WriteSample(r.sample); err != nil {
+				r.framesDropped.Add(1)
+				r.logger.Warn().Err(err).Msg("failed to write sample to WebRTC")
+			} else {
+				r.framesRelayed.Add(1)
+			}
+		}
+	}
+}
+
+// InputRelay forwards audio from WebRTC (browser microphone) to subprocess (USB audio)
+type InputRelay struct {
+	client  *IPCClient
+	ctx     context.Context
+	cancel  context.CancelFunc
+	logger  zerolog.Logger
+	running atomic.Bool
+}
+
+// NewInputRelay creates a relay for input audio (browser → device)
+func NewInputRelay(client *IPCClient) *InputRelay {
+	ctx, cancel := context.WithCancel(context.Background())
+	logger := logging.GetDefaultLogger().With().Str("component", "audio-input-relay").Logger()
+
+	return &InputRelay{
+		client: client,
+		ctx:    ctx,
+		cancel: cancel,
+		logger: logger,
+	}
+}
+
+// Start begins relaying audio frames
+func (r *InputRelay) Start() error {
+	if r.running.Swap(true) {
+		return fmt.Errorf("input relay already running")
+	}
+
+	r.logger.Debug().Msg("input relay started")
+	return nil
+}
+
+// Stop stops the relay
+func (r *InputRelay) Stop() {
+	if !r.running.Swap(false) {
+		return
+	}
+
+	r.cancel()
+	r.logger.Debug().Msg("input relay stopped")
+}
--- a/internal/audio/supervisor.go
+++ b/internal/audio/supervisor.go
@ -0,0 +1,187 @@
+package audio
+
+import (
+	"bufio"
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"sync/atomic"
+	"time"
+
+	"github.com/jetkvm/kvm/internal/logging"
+	"github.com/rs/zerolog"
+)
+
+// Supervisor manages a subprocess lifecycle with automatic restart
+type Supervisor struct {
+	name       string
+	binaryPath string
+	socketPath string
+	env        []string
+
+	cmd     *exec.Cmd
+	ctx     context.Context
+	cancel  context.CancelFunc
+	running atomic.Bool
+	done    chan struct{} // Closed when supervision loop exits
+	logger  zerolog.Logger
+
+	// Restart state
+	restartCount   uint8
+	lastRestartAt  time.Time
+	restartBackoff time.Duration
+}
+
+const (
+	minRestartDelay = 1 * time.Second
+	maxRestartDelay = 30 * time.Second
+	restartWindow   = 5 * time.Minute // Reset backoff if process runs this long
+)
+
+// NewSupervisor creates a new subprocess supervisor
+func NewSupervisor(name, binaryPath, socketPath string, env []string) *Supervisor {
+	ctx, cancel := context.WithCancel(context.Background())
+	logger := logging.GetDefaultLogger().With().Str("component", name).Logger()
+
+	return &Supervisor{
+		name:           name,
+		binaryPath:     binaryPath,
+		socketPath:     socketPath,
+		env:            env,
+		ctx:            ctx,
+		cancel:         cancel,
+		done:           make(chan struct{}),
+		logger:         logger,
+		restartBackoff: minRestartDelay,
+	}
+}
+
+// Start begins supervising the subprocess
+func (s *Supervisor) Start() error {
+	if s.running.Load() {
+		return fmt.Errorf("%s: already running", s.name)
+	}
+
+	s.running.Store(true)
+	go s.supervisionLoop()
+	s.logger.Debug().Msg("supervisor started")
+	return nil
+}
+
+// Stop gracefully stops the subprocess
+func (s *Supervisor) Stop() {
+	if !s.running.Swap(false) {
+		return // Already stopped
+	}
+
+	s.logger.Debug().Msg("stopping supervisor")
+	s.cancel()
+
+	// Kill process if running
+	if s.cmd != nil && s.cmd.Process != nil {
+		_ = s.cmd.Process.Kill() // Ignore error, process may already be dead
+	}
+
+	// Wait for supervision loop to exit
+	<-s.done
+
+	// Clean up socket file
+	os.Remove(s.socketPath)
+	s.logger.Debug().Msg("supervisor stopped")
+}
+
+// supervisionLoop manages the subprocess lifecycle
+func (s *Supervisor) supervisionLoop() {
+	defer close(s.done)
+
+	for s.running.Load() {
+		// Check if we should reset backoff (process ran long enough)
+		if !s.lastRestartAt.IsZero() && time.Since(s.lastRestartAt) > restartWindow {
+			s.restartCount = 0
+			s.restartBackoff = minRestartDelay
+			s.logger.Debug().Msg("reset restart backoff after stable run")
+		}
+
+		// Start the process
+		if err := s.startProcess(); err != nil {
+			s.logger.Error().Err(err).Msg("failed to start process")
+		} else {
+			// Wait for process to exit
+			err := s.cmd.Wait()
+
+			if s.running.Load() {
+				// Process crashed (not intentional shutdown)
+				s.logger.Warn().
+					Err(err).
+					Uint8("restart_count", s.restartCount).
+					Dur("backoff", s.restartBackoff).
+					Msg("process exited unexpectedly, will restart")
+
+				s.restartCount++
+				s.lastRestartAt = time.Now()
+
+				// Calculate next backoff (exponential: 1s, 2s, 4s, 8s, 16s, 30s)
+				s.restartBackoff *= 2
+				if s.restartBackoff > maxRestartDelay {
+					s.restartBackoff = maxRestartDelay
+				}
+
+				// Wait before restart
+				select {
+				case <-time.After(s.restartBackoff):
+					// Continue to next iteration
+				case <-s.ctx.Done():
+					return // Shutting down
+				}
+			} else {
+				// Intentional shutdown
+				s.logger.Debug().Msg("process exited cleanly")
+				return
+			}
+		}
+	}
+}
+
+// logPipe reads from a pipe and logs each line at debug level
+func (s *Supervisor) logPipe(reader io.ReadCloser, stream string) {
+	scanner := bufio.NewScanner(reader)
+	for scanner.Scan() {
+		line := scanner.Text()
+		s.logger.Debug().Str("stream", stream).Msg(line)
+	}
+	reader.Close()
+}
+
+// startProcess starts the subprocess
+func (s *Supervisor) startProcess() error {
+	s.cmd = exec.CommandContext(s.ctx, s.binaryPath)
+	s.cmd.Env = append(os.Environ(), s.env...)
+
+	// Create pipes for subprocess output
+	stdout, err := s.cmd.StdoutPipe()
+	if err != nil {
+		return fmt.Errorf("failed to create stdout pipe: %w", err)
+	}
+	stderr, err := s.cmd.StderrPipe()
+	if err != nil {
+		return fmt.Errorf("failed to create stderr pipe: %w", err)
+	}
+
+	if err := s.cmd.Start(); err != nil {
+		return fmt.Errorf("failed to start %s: %w", s.name, err)
+	}
+
+	// Start goroutines to log subprocess output at debug level
+	go s.logPipe(stdout, "stdout")
+	go s.logPipe(stderr, "stderr")
+
+	s.logger.Debug().
+		Int("pid", s.cmd.Process.Pid).
+		Str("binary", s.binaryPath).
+		Strs("custom_env", s.env).
+		Msg("process started")
+
+	return nil
+}
--- a/internal/usbgadget/changeset_resolver.go
+++ b/internal/usbgadget/changeset_resolver.go
@ -1,7 +1,9 @@
 package usbgadget

 import (
+	"context"
 	"fmt"
+	"time"

 	"github.com/rs/zerolog"
 	"github.com/sourcegraph/tf-dag/dag"
@ -114,7 +116,20 @@ func (c *ChangeSetResolver) resolveChanges(initial bool) error {
 }

 func (c *ChangeSetResolver) applyChanges() error {
+	return c.applyChangesWithTimeout(45 * time.Second)
+}
+
+func (c *ChangeSetResolver) applyChangesWithTimeout(timeout time.Duration) error {
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+
 	for _, change := range c.resolvedChanges {
+		select {
+		case <-ctx.Done():
+			return fmt.Errorf("USB gadget reconfiguration timed out after %v: %w", timeout, ctx.Err())
+		default:
+		}
+
 		change.ResetActionResolution()
 		action := change.Action()
 		actionStr := FileChangeResolvedActionString[action]
@ -126,7 +141,7 @@ func (c *ChangeSetResolver) applyChanges() error {

 		l.Str("action", actionStr).Str("change", change.String()).Msg("applying change")

-		err := c.changeset.applyChange(change)
+		err := c.applyChangeWithTimeout(ctx, change)
 		if err != nil {
 			if change.IgnoreErrors {
 				c.l.Warn().Str("change", change.String()).Err(err).Msg("ignoring error")
@ -139,6 +154,20 @@ func (c *ChangeSetResolver) applyChanges() error {
 	return nil
 }

+func (c *ChangeSetResolver) applyChangeWithTimeout(ctx context.Context, change *FileChange) error {
+	done := make(chan error, 1)
+	go func() {
+		done <- c.changeset.applyChange(change)
+	}()
+
+	select {
+	case err := <-done:
+		return err
+	case <-ctx.Done():
+		return fmt.Errorf("change application timed out for %s: %w", change.String(), ctx.Err())
+	}
+}
+
 func (c *ChangeSetResolver) GetChanges() ([]*FileChange, error) {
 	localChanges := c.changeset.Changes
 	changesMap := make(map[string]*FileChange)
--- a/internal/usbgadget/config.go
+++ b/internal/usbgadget/config.go
@ -59,6 +59,23 @@ var defaultGadgetConfig = map[string]gadgetConfigItem{
 	// mass storage
 	"mass_storage_base": massStorageBaseConfig,
 	"mass_storage_lun0": massStorageLun0Config,
+	// audio (UAC1 - USB Audio Class 1)
+	"audio": {
+		order:      4000,
+		device:     "uac1.usb0",
+		path:       []string{"functions", "uac1.usb0"},
+		configPath: []string{"uac1.usb0"},
+		attrs: gadgetAttributes{
+			"p_chmask":         "3",     // Playback: stereo (2 channels)
+			"p_srate":          "48000", // Playback: 48kHz sample rate
+			"p_ssize":          "2",     // Playback: 16-bit (2 bytes)
+			"p_volume_present": "0",     // Playback: no volume control
+			"c_chmask":         "3",     // Capture: stereo (2 channels)
+			"c_srate":          "48000", // Capture: 48kHz sample rate
+			"c_ssize":          "2",     // Capture: 16-bit (2 bytes)
+			"c_volume_present": "0",     // Capture: no volume control
+		},
+	},
 }

 func (u *UsbGadget) isGadgetConfigItemEnabled(itemKey string) bool {
@ -73,6 +90,8 @@ func (u *UsbGadget) isGadgetConfigItemEnabled(itemKey string) bool {
 		return u.enabledDevices.MassStorage
 	case "mass_storage_lun0":
 		return u.enabledDevices.MassStorage
+	case "audio":
+		return u.enabledDevices.Audio
 	default:
 		return true
 	}
@ -182,6 +201,9 @@ func (u *UsbGadget) Init() error {
 		return u.logError("unable to initialize USB stack", err)
 	}

+	// Pre-open HID files to reduce input latency
+	u.PreOpenHidFiles()
+
 	return nil
 }

@ -191,11 +213,17 @@ func (u *UsbGadget) UpdateGadgetConfig() error {

 	u.loadGadgetConfig()

+	// Close HID files before reconfiguration to prevent "file already closed" errors
+	u.CloseHidFiles()
+
 	err := u.configureUsbGadget(true)
 	if err != nil {
 		return u.logError("unable to update gadget config", err)
 	}

+	// Reopen HID files after reconfiguration
+	u.PreOpenHidFiles()
+
 	return nil
 }

--- a/internal/usbgadget/config_tx.go
+++ b/internal/usbgadget/config_tx.go
@ -1,10 +1,12 @@
 package usbgadget

 import (
+	"context"
 	"fmt"
 	"path"
 	"path/filepath"
 	"sort"
+	"time"

 	"github.com/rs/zerolog"
 )
@ -52,22 +54,49 @@ func (u *UsbGadget) newUsbGadgetTransaction(lock bool) error {
 }

 func (u *UsbGadget) WithTransaction(fn func() error) error {
-	u.txLock.Lock()
-	defer u.txLock.Unlock()
+	return u.WithTransactionTimeout(fn, 60*time.Second)
+}

-	err := u.newUsbGadgetTransaction(false)
-	if err != nil {
-		u.log.Error().Err(err).Msg("failed to create transaction")
-		return err
-	}
-	if err := fn(); err != nil {
-		u.log.Error().Err(err).Msg("transaction failed")
-		return err
-	}
-	result := u.tx.Commit()
-	u.tx = nil
+// WithTransactionTimeout executes a USB gadget transaction with a specified timeout
+// to prevent indefinite blocking during USB reconfiguration operations
+func (u *UsbGadget) WithTransactionTimeout(fn func() error, timeout time.Duration) error {
+	// Create a context with timeout for the entire transaction
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()

-	return result
+	// Channel to signal when the transaction is complete
+	done := make(chan error, 1)
+
+	// Execute the transaction in a goroutine
+	go func() {
+		u.txLock.Lock()
+		defer u.txLock.Unlock()
+
+		err := u.newUsbGadgetTransaction(false)
+		if err != nil {
+			u.log.Error().Err(err).Msg("failed to create transaction")
+			done <- err
+			return
+		}
+
+		if err := fn(); err != nil {
+			u.log.Error().Err(err).Msg("transaction failed")
+			done <- err
+			return
+		}
+
+		result := u.tx.Commit()
+		u.tx = nil
+		done <- result
+	}()
+
+	// Wait for either completion or timeout
+	select {
+	case err := <-done:
+		return err
+	case <-ctx.Done():
+		return fmt.Errorf("USB gadget transaction timed out after %v: %w", timeout, ctx.Err())
+	}
 }

 func (tx *UsbGadgetTransaction) addFileChange(component string, change RequestedFileChange) string {
--- a/internal/usbgadget/usbgadget.go
+++ b/internal/usbgadget/usbgadget.go
@ -19,6 +19,7 @@ type Devices struct {
 	RelativeMouse bool `json:"relative_mouse"`
 	Keyboard      bool `json:"keyboard"`
 	MassStorage   bool `json:"mass_storage"`
+	Audio         bool `json:"audio"`
 }

 // Config is a struct that represents the customizations for a USB gadget.
@ -39,6 +40,7 @@ var defaultUsbGadgetDevices = Devices{
 	RelativeMouse: true,
 	Keyboard:      true,
 	MassStorage:   true,
+	Audio:         true,
 }

 type KeysDownState struct {
@ -188,3 +190,63 @@ func (u *UsbGadget) Close() error {

 	return nil
 }
+
+// CloseHidFiles closes all open HID files
+func (u *UsbGadget) CloseHidFiles() {
+	u.log.Debug().Msg("closing HID files")
+
+	// Close keyboard HID file
+	if u.keyboardHidFile != nil {
+		if err := u.keyboardHidFile.Close(); err != nil {
+			u.log.Debug().Err(err).Msg("failed to close keyboard HID file")
+		}
+		u.keyboardHidFile = nil
+	}
+
+	// Close absolute mouse HID file
+	if u.absMouseHidFile != nil {
+		if err := u.absMouseHidFile.Close(); err != nil {
+			u.log.Debug().Err(err).Msg("failed to close absolute mouse HID file")
+		}
+		u.absMouseHidFile = nil
+	}
+
+	// Close relative mouse HID file
+	if u.relMouseHidFile != nil {
+		if err := u.relMouseHidFile.Close(); err != nil {
+			u.log.Debug().Err(err).Msg("failed to close relative mouse HID file")
+		}
+		u.relMouseHidFile = nil
+	}
+}
+
+// PreOpenHidFiles opens all HID files to reduce input latency
+func (u *UsbGadget) PreOpenHidFiles() {
+	// Add a small delay to allow USB gadget reconfiguration to complete
+	// This prevents "no such device or address" errors when trying to open HID files
+	time.Sleep(100 * time.Millisecond)
+
+	if u.enabledDevices.Keyboard {
+		if err := u.openKeyboardHidFile(); err != nil {
+			u.log.Debug().Err(err).Msg("failed to pre-open keyboard HID file")
+		}
+	}
+	if u.enabledDevices.AbsoluteMouse {
+		if u.absMouseHidFile == nil {
+			var err error
+			u.absMouseHidFile, err = os.OpenFile("/dev/hidg1", os.O_RDWR, 0666)
+			if err != nil {
+				u.log.Debug().Err(err).Msg("failed to pre-open absolute mouse HID file")
+			}
+		}
+	}
+	if u.enabledDevices.RelativeMouse {
+		if u.relMouseHidFile == nil {
+			var err error
+			u.relMouseHidFile, err = os.OpenFile("/dev/hidg2", os.O_RDWR, 0666)
+			if err != nil {
+				u.log.Debug().Err(err).Msg("failed to pre-open relative mouse HID file")
+			}
+		}
+	}
+}
--- a/jsonrpc.go
+++ b/jsonrpc.go
@ -700,7 +700,8 @@ func rpcSetUsbConfig(usbConfig usbgadget.Config) error {
 	LoadConfig()
 	config.UsbConfig = &usbConfig
 	gadget.SetGadgetConfig(config.UsbConfig)
-	return updateUsbRelatedConfig()
+	wasAudioEnabled := config.UsbDevices != nil && config.UsbDevices.Audio
+	return updateUsbRelatedConfig(wasAudioEnabled)
 }

 func rpcGetWakeOnLanDevices() ([]WakeOnLanDevice, error) {
@ -911,23 +912,67 @@ func rpcGetUsbDevices() (usbgadget.Devices, error) {
 	return *config.UsbDevices, nil
 }

-func updateUsbRelatedConfig() error {
+func updateUsbRelatedConfig(wasAudioEnabled bool) error {
+	ensureConfigLoaded()
+
+	audioSourceChanged := false
+
+	// If USB audio is being disabled and audio output source is USB, switch to HDMI
+	if config.UsbDevices != nil && !config.UsbDevices.Audio && config.AudioOutputSource == "usb" {
+		audioMutex.Lock()
+		config.AudioOutputSource = "hdmi"
+		useUSBForAudioOutput = false
+		audioSourceChanged = true
+		audioMutex.Unlock()
+	}
+
+	// If USB audio is being enabled (was disabled, now enabled), switch to USB
+	if config.UsbDevices != nil && config.UsbDevices.Audio && !wasAudioEnabled {
+		audioMutex.Lock()
+		config.AudioOutputSource = "usb"
+		useUSBForAudioOutput = true
+		audioSourceChanged = true
+		audioMutex.Unlock()
+	}
+
+	// Stop audio subprocesses before USB reconfiguration
+	// Input always uses USB, output depends on audioSourceChanged
+	audioMutex.Lock()
+	stopInputSubprocessLocked()
+	if audioSourceChanged {
+		stopOutputSubprocessLocked()
+	}
+	audioMutex.Unlock()
+
 	if err := gadget.UpdateGadgetConfig(); err != nil {
 		return fmt.Errorf("failed to write gadget config: %w", err)
 	}
+
 	if err := SaveConfig(); err != nil {
 		return fmt.Errorf("failed to save config: %w", err)
 	}
+
+	// Restart audio if source changed or USB audio is enabled with active connections
+	// The subprocess supervisor and relay handle device readiness via retry logic
+	if activeConnections.Load() > 0 && (audioSourceChanged || (config.UsbDevices != nil && config.UsbDevices.Audio)) {
+		if err := startAudioSubprocesses(); err != nil {
+			logger.Warn().Err(err).Msg("Failed to restart audio after USB reconfiguration")
+		}
+	}
+
 	return nil
 }

 func rpcSetUsbDevices(usbDevices usbgadget.Devices) error {
+	wasAudioEnabled := config.UsbDevices != nil && config.UsbDevices.Audio
 	config.UsbDevices = &usbDevices
 	gadget.SetGadgetDevices(config.UsbDevices)
-	return updateUsbRelatedConfig()
+	return updateUsbRelatedConfig(wasAudioEnabled)
 }

 func rpcSetUsbDeviceState(device string, enabled bool) error {
+	wasAudioEnabled := config.UsbDevices != nil && config.UsbDevices.Audio
+
 	switch device {
 	case "absoluteMouse":
 		config.UsbDevices.AbsoluteMouse = enabled
@ -937,11 +982,43 @@ func rpcSetUsbDeviceState(device string, enabled bool) error {
 		config.UsbDevices.Keyboard = enabled
 	case "massStorage":
 		config.UsbDevices.MassStorage = enabled
+	case "audio":
+		config.UsbDevices.Audio = enabled
 	default:
 		return fmt.Errorf("invalid device: %s", device)
 	}
 	gadget.SetGadgetDevices(config.UsbDevices)
-	return updateUsbRelatedConfig()
+	return updateUsbRelatedConfig(wasAudioEnabled)
+}
+
+func rpcGetAudioOutputSource() (string, error) {
+	ensureConfigLoaded()
+	return config.AudioOutputSource, nil
+}
+
+func rpcSetAudioOutputSource(source string) error {
+	if source != "hdmi" && source != "usb" {
+		return fmt.Errorf("invalid audio output source: %s (must be 'hdmi' or 'usb')", source)
+	}
+
+	useUSB := source == "usb"
+	return SetAudioOutputSource(useUSB)
+}
+
+func rpcGetAudioOutputEnabled() (bool, error) {
+	return audioOutputEnabled.Load(), nil
+}
+
+func rpcSetAudioOutputEnabled(enabled bool) error {
+	return SetAudioOutputEnabled(enabled)
+}
+
+func rpcGetAudioInputEnabled() (bool, error) {
+	return audioInputEnabled.Load(), nil
+}
+
+func rpcSetAudioInputEnabled(enabled bool) error {
+	return SetAudioInputEnabled(enabled)
 }

 func rpcSetCloudUrl(apiUrl string, appUrl string) error {
@ -1260,6 +1337,12 @@ var rpcHandlers = map[string]RPCHandler{
 	"getUsbDevices":          {Func: rpcGetUsbDevices},
 	"setUsbDevices":          {Func: rpcSetUsbDevices, Params: []string{"devices"}},
 	"setUsbDeviceState":      {Func: rpcSetUsbDeviceState, Params: []string{"device", "enabled"}},
+	"getAudioOutputSource":   {Func: rpcGetAudioOutputSource},
+	"setAudioOutputSource":   {Func: rpcSetAudioOutputSource, Params: []string{"source"}},
+	"getAudioOutputEnabled":  {Func: rpcGetAudioOutputEnabled},
+	"setAudioOutputEnabled":  {Func: rpcSetAudioOutputEnabled, Params: []string{"enabled"}},
+	"getAudioInputEnabled":   {Func: rpcGetAudioInputEnabled},
+	"setAudioInputEnabled":   {Func: rpcSetAudioInputEnabled, Params: []string{"enabled"}},
 	"setCloudUrl":            {Func: rpcSetCloudUrl, Params: []string{"apiUrl", "appUrl"}},
 	"getKeyboardLayout":      {Func: rpcGetKeyboardLayout},
 	"setKeyboardLayout":      {Func: rpcSetKeyboardLayout, Params: []string{"layout"}},
--- a/main.go
+++ b/main.go
@ -34,6 +34,7 @@ func Main() {
 	go confirmCurrentSystem()

 	initNative(systemVersionLocal, appVersionLocal)
+	initAudio()

 	http.DefaultClient.Timeout = 1 * time.Minute

@ -123,6 +124,9 @@ func Main() {
 	signal.Notify(sigs, syscall.SIGINT, syscall.SIGTERM)
 	<-sigs
 	logger.Info().Msg("JetKVM Shutting Down")
+
+	stopAudioSubprocesses()
+
 	//if fuseServer != nil {
 	//	err := setMassStorageImage(" ")
 	//	if err != nil {
--- a/ui/src/components/ActionBar.tsx
+++ b/ui/src/components/ActionBar.tsx
@ -1,5 +1,5 @@
 import { MdOutlineContentPasteGo } from "react-icons/md";
-import { LuCable, LuHardDrive, LuMaximize, LuSettings, LuSignal } from "react-icons/lu";
+import { LuCable, LuHardDrive, LuMaximize, LuSettings, LuSignal, LuVolume2 } from "react-icons/lu";
 import { FaKeyboard } from "react-icons/fa6";
 import { Popover, PopoverButton, PopoverPanel } from "@headlessui/react";
 import { Fragment, useCallback, useRef } from "react";
@ -18,6 +18,7 @@ import PasteModal from "@/components/popovers/PasteModal";
 import WakeOnLanModal from "@/components/popovers/WakeOnLan/Index";
 import MountPopopover from "@/components/popovers/MountPopover";
 import ExtensionPopover from "@/components/popovers/ExtensionPopover";
+import AudioPopover from "@/components/popovers/AudioPopover";
 import { useDeviceUiNavigation } from "@/hooks/useAppNavigation";

 export default function Actionbar({
@ -203,6 +204,36 @@ export default function Actionbar({
              onClick={() => setVirtualKeyboardEnabled(!isVirtualKeyboardEnabled)}
            />
          </div>
+          <Popover>
+            <PopoverButton as={Fragment}>
+              <Button
+                size="XS"
+                theme="light"
+                text="Audio"
+                LeadingIcon={LuVolume2}
+                onClick={() => {
+                  setDisableVideoFocusTrap(true);
+                }}
+              />
+            </PopoverButton>
+            <PopoverPanel
+              anchor="bottom start"
+              transition
+              className={cx(
+                "z-10 flex w-[420px] flex-col overflow-visible!",
+                "flex origin-top flex-col transition duration-300 ease-out data-closed:translate-y-8 data-closed:opacity-0",
+              )}
+            >
+              {({ open }) => {
+                checkIfStateChanged(open);
+                return (
+                  <div className="mx-auto w-full max-w-xl">
+                    <AudioPopover />
+                  </div>
+                );
+              }}
+            </PopoverPanel>
+          </Popover>
        </div>

        <div className="flex flex-wrap items-center gap-x-2 gap-y-2">
--- a/ui/src/components/UsbDeviceSetting.tsx
+++ b/ui/src/components/UsbDeviceSetting.tsx
@ -23,6 +23,7 @@ export interface UsbDeviceConfig {
  absolute_mouse: boolean;
  relative_mouse: boolean;
  mass_storage: boolean;
+  audio: boolean;
 }

 const defaultUsbDeviceConfig: UsbDeviceConfig = {
@ -30,17 +31,19 @@ const defaultUsbDeviceConfig: UsbDeviceConfig = {
  absolute_mouse: true,
  relative_mouse: true,
  mass_storage: true,
+  audio: true,
 };

 const usbPresets = [
  {
-    label: "Keyboard, Mouse and Mass Storage",
+    label: "Keyboard, Mouse, Mass Storage and Audio",
    value: "default",
    config: {
      keyboard: true,
      absolute_mouse: true,
      relative_mouse: true,
      mass_storage: true,
+      audio: true,
    },
  },
  {
@ -51,6 +54,7 @@ const usbPresets = [
      absolute_mouse: false,
      relative_mouse: false,
      mass_storage: false,
+      audio: false,
    },
  },
  {
@ -218,6 +222,17 @@ export function UsbDeviceSetting() {
                />
              </SettingsItem>
            </div>
+            <div className="space-y-4">
+              <SettingsItem
+                title="Enable USB Audio"
+                description="Enable bidirectional audio (HDMI capture and microphone input)"
+              >
+                <Checkbox
+                  checked={usbDeviceConfig.audio}
+                  onChange={onUsbConfigItemChange("audio")}
+                />
+              </SettingsItem>
+            </div>
          </div>
          <div className="mt-6 flex gap-x-2">
            <Button
--- a/ui/src/components/WebRTCVideo.tsx
+++ b/ui/src/components/WebRTCVideo.tsx
@ -318,17 +318,31 @@ export default function WebRTCVideo() {
      if (!peerConnection) return;
      const abortController = new AbortController();
      const signal = abortController.signal;
+      const audioElements: HTMLAudioElement[] = [];

      peerConnection.addEventListener(
        "track",
        (e: RTCTrackEvent) => {
-          addStreamToVideoElm(e.streams[0]);
+          if (e.track.kind === "video") {
+            addStreamToVideoElm(e.streams[0]);
+          } else if (e.track.kind === "audio") {
+            const audioElm = document.createElement("audio");
+            audioElm.autoplay = true;
+            audioElm.srcObject = e.streams[0];
+            audioElm.style.display = "none";
+            document.body.appendChild(audioElm);
+            audioElements.push(audioElm);
+          }
        },
        { signal },
      );

      return () => {
        abortController.abort();
+        audioElements.forEach((audioElm) => {
+          audioElm.srcObject = null;
+          audioElm.remove();
+        });
      };
    },
    [addStreamToVideoElm, peerConnection],
@ -518,7 +532,6 @@ export default function WebRTCVideo() {
                        controls={false}
                        onPlaying={onVideoPlaying}
                        onPlay={onVideoPlaying}
-                        muted
                        playsInline
                        disablePictureInPicture
                        controlsList="nofullscreen"
--- a/ui/src/components/popovers/AudioPopover.tsx
+++ b/ui/src/components/popovers/AudioPopover.tsx
@ -0,0 +1,182 @@
+import { useCallback, useEffect, useState } from "react";
+import { LuVolume2 } from "react-icons/lu";
+
+import { JsonRpcResponse, useJsonRpc } from "@/hooks/useJsonRpc";
+import { GridCard } from "@components/Card";
+import { SettingsItem } from "@components/SettingsItem";
+import { SelectMenuBasic } from "@components/SelectMenuBasic";
+import { Button } from "@components/Button";
+import notifications from "@/notifications";
+
+export default function AudioPopover() {
+  const { send } = useJsonRpc();
+  const [audioOutputSource, setAudioOutputSource] = useState<string>("hdmi");
+  const [audioOutputEnabled, setAudioOutputEnabled] = useState<boolean>(true);
+  const [audioInputEnabled, setAudioInputEnabled] = useState<boolean>(true);
+  const [usbAudioEnabled, setUsbAudioEnabled] = useState<boolean>(false);
+  const [loading, setLoading] = useState(false);
+
+  useEffect(() => {
+    // Load current audio settings
+    send("getAudioOutputSource", {}, (resp: JsonRpcResponse) => {
+      if ("error" in resp) {
+        console.error("Failed to load audio output source:", resp.error);
+      } else {
+        setAudioOutputSource(resp.result as string);
+      }
+    });
+
+    send("getAudioOutputEnabled", {}, (resp: JsonRpcResponse) => {
+      if ("error" in resp) {
+        console.error("Failed to load audio output enabled:", resp.error);
+      } else {
+        setAudioOutputEnabled(resp.result as boolean);
+      }
+    });
+
+    send("getAudioInputEnabled", {}, (resp: JsonRpcResponse) => {
+      if ("error" in resp) {
+        console.error("Failed to load audio input enabled:", resp.error);
+      } else {
+        setAudioInputEnabled(resp.result as boolean);
+      }
+    });
+
+    send("getUsbDevices", {}, (resp: JsonRpcResponse) => {
+      if ("error" in resp) {
+        console.error("Failed to load USB devices:", resp.error);
+      } else {
+        const usbDevices = resp.result as { audio: boolean };
+        setUsbAudioEnabled(usbDevices.audio || false);
+      }
+    });
+  }, [send]);
+
+  const handleAudioOutputSourceChange = useCallback(
+    (e: React.ChangeEvent<HTMLSelectElement>) => {
+      const newSource = e.target.value;
+      setLoading(true);
+      send("setAudioOutputSource", { source: newSource }, (resp: JsonRpcResponse) => {
+        setLoading(false);
+        if ("error" in resp) {
+          notifications.error(
+            `Failed to set audio output source: ${resp.error.data || "Unknown error"}`,
+          );
+        } else {
+          setAudioOutputSource(newSource);
+          notifications.success(`Audio output source set to ${newSource.toUpperCase()}`);
+        }
+      });
+    },
+    [send],
+  );
+
+  const handleAudioOutputEnabledToggle = useCallback(
+    (e: React.ChangeEvent<HTMLInputElement>) => {
+      const enabled = e.target.checked;
+      setLoading(true);
+      send("setAudioOutputEnabled", { enabled }, (resp: JsonRpcResponse) => {
+        setLoading(false);
+        if ("error" in resp) {
+          notifications.error(
+            `Failed to ${enabled ? "enable" : "disable"} audio output: ${resp.error.data || "Unknown error"}`,
+          );
+        } else {
+          setAudioOutputEnabled(enabled);
+          notifications.success(`Audio output ${enabled ? "enabled" : "disabled"}`);
+        }
+      });
+    },
+    [send],
+  );
+
+  const handleAudioInputEnabledToggle = useCallback(
+    (e: React.ChangeEvent<HTMLInputElement>) => {
+      const enabled = e.target.checked;
+      setLoading(true);
+      send("setAudioInputEnabled", { enabled }, (resp: JsonRpcResponse) => {
+        setLoading(false);
+        if ("error" in resp) {
+          notifications.error(
+            `Failed to ${enabled ? "enable" : "disable"} audio input: ${resp.error.data || "Unknown error"}`,
+          );
+        } else {
+          setAudioInputEnabled(enabled);
+          notifications.success(`Audio input ${enabled ? "enabled" : "disabled"}`);
+        }
+      });
+    },
+    [send],
+  );
+
+  return (
+    <GridCard>
+      <div className="space-y-4 p-4 py-3">
+        <div className="space-y-4">
+          <div className="flex items-center gap-2 text-slate-900 dark:text-slate-100">
+            <LuVolume2 className="h-5 w-5" />
+            <h3 className="font-semibold">Audio Settings</h3>
+          </div>
+
+          <div className="space-y-3">
+            <SettingsItem
+              loading={loading}
+              title="Audio Output"
+              description="Enable audio from target to speakers"
+            >
+              <Button
+                size="SM"
+                theme={audioOutputEnabled ? "light" : "primary"}
+                text={audioOutputEnabled ? "Disable" : "Enable"}
+                onClick={() => handleAudioOutputEnabledToggle({ target: { checked: !audioOutputEnabled } } as React.ChangeEvent<HTMLInputElement>)}
+              />
+            </SettingsItem>
+
+            <SettingsItem
+              loading={loading}
+              title="Audio Output Source"
+              description={usbAudioEnabled ? "Select where to capture audio from" : "Enable USB Audio to use USB as source"}
+            >
+              <SelectMenuBasic
+                size="SM"
+                label=""
+                className="max-w-[180px]"
+                value={audioOutputSource}
+                fullWidth
+                disabled={!audioOutputEnabled}
+                onChange={handleAudioOutputSourceChange}
+                options={
+                  usbAudioEnabled
+                    ? [
+                        { label: "HDMI", value: "hdmi" },
+                        { label: "USB", value: "usb" },
+                      ]
+                    : [{ label: "HDMI", value: "hdmi" }]
+                }
+              />
+            </SettingsItem>
+
+            {usbAudioEnabled && (
+              <>
+                <div className="h-px w-full bg-slate-800/10 dark:bg-slate-300/20" />
+
+                <SettingsItem
+                  loading={loading}
+                  title="Audio Input (Microphone)"
+                  description="Enable microphone input to target"
+                >
+                  <Button
+                    size="SM"
+                    theme={audioInputEnabled ? "light" : "primary"}
+                    text={audioInputEnabled ? "Disable" : "Enable"}
+                    onClick={() => handleAudioInputEnabledToggle({ target: { checked: !audioInputEnabled } } as React.ChangeEvent<HTMLInputElement>)}
+                  />
+                </SettingsItem>
+              </>
+            )}
+          </div>
+        </div>
+      </div>
+    </GridCard>
+  );
+}
--- a/ui/src/routes/devices.$id.tsx
+++ b/ui/src/routes/devices.$id.tsx
@ -184,6 +184,31 @@ export default function KvmIdRoute() {
    ) {
      setLoadingMessage("Setting remote description");

+      // Enable stereo in remote answer SDP
+      if (remoteDescription.sdp) {
+        const opusMatch = remoteDescription.sdp.match(/a=rtpmap:(\d+)\s+opus\/48000\/2/i);
+        if (!opusMatch) {
+          console.warn("[SDP] Opus 48kHz stereo not found in answer - stereo may not work");
+        } else {
+          const pt = opusMatch[1];
+          const stereoParams = 'stereo=1;sprop-stereo=1;maxaveragebitrate=128000';
+          const fmtpRegex = new RegExp(`a=fmtp:${pt}\\s+(.+)`, 'i');
+          const fmtpMatch = remoteDescription.sdp.match(fmtpRegex);
+
+          if (fmtpMatch && !fmtpMatch[1].includes('stereo=')) {
+            remoteDescription.sdp = remoteDescription.sdp.replace(
+              fmtpRegex,
+              `a=fmtp:${pt} ${fmtpMatch[1]};${stereoParams}`
+            );
+          } else if (!fmtpMatch) {
+            remoteDescription.sdp = remoteDescription.sdp.replace(
+              opusMatch[0],
+              `${opusMatch[0]}\r\na=fmtp:${pt} ${stereoParams}`
+            );
+          }
+        }
+      }
+
      try {
        await pc.setRemoteDescription(new RTCSessionDescription(remoteDescription));
        console.log("[setRemoteSessionDescription] Remote description set successfully");
@ -430,6 +455,30 @@ export default function KvmIdRoute() {
        makingOffer.current = true;

        const offer = await pc.createOffer();
+
+        // Enable stereo for Opus audio codec
+        if (offer.sdp) {
+          const opusMatch = offer.sdp.match(/a=rtpmap:(\d+)\s+opus\/48000\/2/i);
+          if (!opusMatch) {
+            console.warn("[SDP] Opus 48kHz stereo not found in offer - stereo may not work");
+          } else {
+            const pt = opusMatch[1];
+            const stereoParams = 'stereo=1;sprop-stereo=1;maxaveragebitrate=128000';
+            const fmtpRegex = new RegExp(`a=fmtp:${pt}\\s+(.+)`, 'i');
+            const fmtpMatch = offer.sdp.match(fmtpRegex);
+
+            if (fmtpMatch) {
+              // Modify existing fmtp line
+              if (!fmtpMatch[1].includes('stereo=')) {
+                offer.sdp = offer.sdp.replace(fmtpRegex, `a=fmtp:${pt} ${fmtpMatch[1]};${stereoParams}`);
+              }
+            } else {
+              // Add new fmtp line after rtpmap
+              offer.sdp = offer.sdp.replace(opusMatch[0], `${opusMatch[0]}\r\na=fmtp:${pt} ${stereoParams}`);
+            }
+          }
+        }
+
        await pc.setLocalDescription(offer);
        const sd = btoa(JSON.stringify(pc.localDescription));
        const isNewSignalingEnabled = isLegacySignalingEnabled.current === false;
@ -472,11 +521,32 @@ export default function KvmIdRoute() {
    };

    pc.ontrack = function (event) {
-      setMediaStream(event.streams[0]);
+      if (event.track.kind === "video") {
+        setMediaStream(event.streams[0]);
+      }
    };

    setTransceiver(pc.addTransceiver("video", { direction: "recvonly" }));

+    const audioTransceiver = pc.addTransceiver("audio", { direction: "sendrecv" });
+
+    navigator.mediaDevices.getUserMedia({
+      audio: {
+        echoCancellation: true,
+        noiseSuppression: true,
+        autoGainControl: true,
+        channelCount: 2, // Request stereo input if available
+      }
+    }).then((stream) => {
+      const audioTrack = stream.getAudioTracks()[0];
+      if (audioTrack && audioTransceiver.sender) {
+        audioTransceiver.sender.replaceTrack(audioTrack);
+        console.debug("[setupPeerConnection] Audio track settings:", audioTrack.getSettings());
+      }
+    }).catch((err) => {
+      console.warn("Microphone access denied or unavailable:", err.message);
+    });
+
    const rpcDataChannel = pc.createDataChannel("rpc");
    rpcDataChannel.onopen = () => {
      setRpcDataChannel(rpcDataChannel);
--- a/webrtc.go
+++ b/webrtc.go
@ -22,6 +22,7 @@ import (
 type Session struct {
 	peerConnection           *webrtc.PeerConnection
 	VideoTrack               *webrtc.TrackLocalStaticSample
+	AudioTrack               *webrtc.TrackLocalStaticSample
 	ControlChannel           *webrtc.DataChannel
 	RPCChannel               *webrtc.DataChannel
 	HidChannel               *webrtc.DataChannel
@ -295,6 +296,39 @@ func newSession(config SessionConfig) (*Session, error) {
 			}
 		}
 	}()
+
+	session.AudioTrack, err = webrtc.NewTrackLocalStaticSample(
+		webrtc.RTPCodecCapability{MimeType: webrtc.MimeTypeOpus},
+		"audio",
+		"kvm-audio",
+	)
+	if err != nil {
+		scopedLogger.Warn().Err(err).Msg("Failed to create AudioTrack (non-fatal)")
+	} else {
+		_, err = peerConnection.AddTransceiverFromTrack(session.AudioTrack, webrtc.RTPTransceiverInit{
+			Direction: webrtc.RTPTransceiverDirectionSendrecv,
+		})
+		if err != nil {
+			scopedLogger.Warn().Err(err).Msg("Failed to add AudioTrack transceiver (non-fatal)")
+			session.AudioTrack = nil
+		} else {
+			setAudioTrack(session.AudioTrack)
+
+			peerConnection.OnTrack(func(track *webrtc.TrackRemote, receiver *webrtc.RTPReceiver) {
+				scopedLogger.Info().
+					Str("codec", track.Codec().MimeType).
+					Str("track_id", track.ID()).
+					Msg("Received incoming audio track from browser")
+
+				// Store track for connection when audio subprocesses start
+				// OnTrack fires during SDP exchange, before ICE connection completes
+				setPendingInputTrack(track)
+			})
+
+			scopedLogger.Info().Msg("Audio tracks configured successfully")
+		}
+	}
+
 	var isConnected bool

 	peerConnection.OnICECandidate(func(candidate *webrtc.ICECandidate) {
@ -326,6 +360,8 @@ func newSession(config SessionConfig) (*Session, error) {
 		}
 		if connectionState == webrtc.ICEConnectionStateClosed {
 			scopedLogger.Debug().Msg("ICE Connection State is closed, unmounting virtual media")
+			// Only clear currentSession if this is actually the current session
+			// This prevents race condition where old session closes after new one connects
 			if session == currentSession {
 				// Cancel any ongoing keyboard report multi when session closes
 				cancelKeyboardMacro()
@ -372,8 +408,10 @@ func onActiveSessionsChanged() {

 func onFirstSessionConnected() {
 	_ = nativeInstance.VideoStart()
+	onWebRTCConnect()
 }

 func onLastSessionDisconnected() {
 	_ = nativeInstance.VideoStop()
+	onWebRTCDisconnect()
 }