@elizaos/capacitor-talkmode 2.0.0-beta.1 → 2.0.3-beta.3

package/android/src/main/java/ai/eliza/plugins/talkmode/TalkModePlugin.kt

@@ -2,10 +2,17 @@ package ai.eliza.plugins.talkmode
 
 import android.Manifest
 import android.content.Intent
+import android.content.pm.PackageManager
 import android.media.AudioAttributes
+import android.media.AudioDeviceInfo
+import android.media.AudioFocusRequest
 import android.media.AudioFormat
 import android.media.AudioManager
+import android.media.AudioRecord
 import android.media.AudioTrack
+import android.media.MediaRecorder
+import android.util.Base64
+import android.os.Build
 import android.os.Bundle
 import android.os.Handler
 import android.os.Looper
@@ -25,7 +32,15 @@ import com.getcapacitor.annotation.CapacitorPlugin
 import com.getcapacitor.annotation.Permission
 import com.getcapacitor.annotation.PermissionCallback
 import kotlinx.coroutines.*
+import android.net.LocalSocket
+import android.net.LocalSocketAddress
 import java.io.BufferedInputStream
+import java.io.ByteArrayInputStream
+import java.io.DataInputStream
+import java.io.DataOutputStream
+import java.nio.ByteBuffer
+import java.nio.ByteOrder
+import java.io.File
 import java.net.HttpURLConnection
 import java.net.URL
 import java.util.Locale
@@ -45,6 +60,15 @@ class TalkModePlugin : Plugin() {
         private const val TAG = "TalkMode"
         private const val DEFAULT_MODEL_ID = "eleven_flash_v2_5"
         private const val DEFAULT_OUTPUT_FORMAT = "pcm_24000"
+        private const val LOCAL_INFERENCE_TTS_URL = "http://127.0.0.1:31337/api/tts/local-inference"
+        // Abstract-namespace UDS of ElizaBionicInferenceServer (the bionic app
+        // process that has libelizainference loaded). Kept in sync with
+        // BIONIC_INFERENCE_SOCKET_NAME in ElizaAgentService.
+        private const val BIONIC_INFER_SOCKET = "eliza_bionic_infer_v1"
+        // 16 kHz mono is the rate VAD / diarizer / wake-word models expect; 20 ms
+        // (320 samples) is the standard VAD frame size.
+        private const val DEFAULT_FRAME_SAMPLE_RATE = 16000
+        private const val DEFAULT_FRAME_MS = 20
     }
 
     private val mainHandler = Handler(Looper.getMainLooper())
@@ -60,11 +84,19 @@ class TalkModePlugin : Plugin() {
     private var isListening = false
     private var listeningMode = false
     private var stopRequested = false
+    // Consecutive ERROR_NO_MATCH/SPEECH_TIMEOUT count, for exponential restart
+    // backoff so an idle always-on session settles instead of re-arming (and,
+    // with the system recognizer, beeping) every ~600ms when nobody is talking.
+    private var consecutiveNoMatch = 0
     private var restartJob: Job? = null
     private var lastTranscript = ""
     private var lastHeardAtMs: Long? = null
     private var silenceJob: Job? = null
     private val silenceWindowMs = 700L
+    // The recognizer's own onResults AND our silence monitor can both finalize
+    // the same utterance; dedup so a turn is emitted (and sent) exactly once.
+    private var lastEmittedFinal = ""
+    private var lastEmittedFinalAtMs = 0L
 
     // TTS
     private var systemTts: TextToSpeech? = null
@@ -79,10 +111,37 @@ class TalkModePlugin : Plugin() {
     private var lastSpokenText: String? = null
     private var speakStartTimeMs: Long = 0
     private var lastInterruptedAtSeconds: Double? = null
+    @Volatile private var activePcmConnection: HttpURLConnection? = null
 
-    // Audio focus
+    // Voice audio session (communication-mode routing + focus, mirrors the iOS
+    // .playAndRecord/.voiceChat/.defaultToSpeaker session). Held for the whole
+    // conversation so the platform AEC has a stable speaker reference to cancel.
     private var audioManager: AudioManager? = null
-    private var audioFocusRequest: AudioManager.OnAudioFocusChangeListener? = null
+    private var audioFocusRequest: AudioFocusRequest? = null
+    private var audioSessionActive = false
+    private var savedAudioMode = AudioManager.MODE_NORMAL
+    private var savedSpeakerphoneOn = false
+    // Streams we mute for the session to suppress the platform recognizer's
+    // start/stop earcons (the "on/off" beeps heard as it re-arms continuously).
+    // TTS plays on STREAM_VOICE_CALL (USAGE_VOICE_COMMUNICATION) so it stays
+    // audible. Tracked so we only unmute streams we muted.
+    private val earconStreams = intArrayOf(
+        AudioManager.STREAM_MUSIC,
+        AudioManager.STREAM_SYSTEM,
+        AudioManager.STREAM_NOTIFICATION,
+    )
+    private var earconStreamsMuted = false
+
+    // Raw PCM frame capture (diarization / VAD / wake-word source). Opt-in and
+    // mutually exclusive with SpeechRecognizer on the mic: Android only lets one
+    // capture client own a given input source at a time, so starting frame
+    // capture SUSPENDS any active SpeechRecognizer and stopping it resumes STT.
+    private var audioRecord: AudioRecord? = null
+    private var audioFrameJob: Job? = null
+    private val audioFrameRunning = AtomicBoolean(false)
+    private var sttSuspendedForFrames = false
+    private var lastFrameSampleRate = DEFAULT_FRAME_SAMPLE_RATE
+    private var lastFrameSamples = 0
 
     // Config
     private var apiKey: String? = null
@@ -106,6 +165,7 @@ class TalkModePlugin : Plugin() {
 
         override fun onBeginningOfSpeech() {
             Log.d(TAG, "Beginning of speech")
+            consecutiveNoMatch = 0
         }
 
         override fun onRmsChanged(rmsdB: Float) {}
@@ -142,24 +202,34 @@ class TalkModePlugin : Plugin() {
                 return
             }
 
-            // Don't notify error for no-match / speech-timeout, just restart
-            if (error != SpeechRecognizer.ERROR_NO_MATCH &&
-                error != SpeechRecognizer.ERROR_SPEECH_TIMEOUT
+            // Don't notify error for no-match / speech-timeout, just restart.
+            // These fire continuously when the always-on session hears only
+            // silence, so back off exponentially (600ms → 8s cap) instead of
+            // re-arming the recognizer every 600ms. onBeginningOfSpeech /
+            // onResults reset the counter the moment real speech arrives.
+            if (error == SpeechRecognizer.ERROR_NO_MATCH ||
+                error == SpeechRecognizer.ERROR_SPEECH_TIMEOUT
             ) {
+                consecutiveNoMatch++
+                scheduleRestart(
+                    delayMs = minOf(600L * (1L shl minOf(consecutiveNoMatch, 4)), 8000L),
+                )
+            } else {
+                consecutiveNoMatch = 0
                 notifyListeners("error", JSObject().apply {
                     put("code", "recognition_error")
                     put("message", errorMsg)
                     put("recoverable", true)
                 })
+                scheduleRestart(delayMs = 600)
             }
-
-            scheduleRestart(delayMs = 600)
         }
 
         override fun onResults(results: Bundle?) {
             val matches = results?.getStringArrayList(SpeechRecognizer.RESULTS_RECOGNITION)
             val transcript = matches?.firstOrNull()?.trim() ?: ""
             if (transcript.isNotEmpty()) {
+                consecutiveNoMatch = 0
                 handleTranscript(transcript, isFinal = true)
             }
             scheduleRestart()
@@ -189,6 +259,7 @@ class TalkModePlugin : Plugin() {
             systemTtsReady = status == TextToSpeech.SUCCESS
             if (systemTtsReady) {
                 systemTts?.language = Locale.getDefault()
+                systemTts?.setAudioAttributes(voiceAudioAttributes())
                 systemTts?.setOnUtteranceProgressListener(object : UtteranceProgressListener() {
                     override fun onStart(id: String?) {}
 
@@ -270,14 +341,13 @@ class TalkModePlugin : Plugin() {
         enabled = true
         stopRequested = false
         listeningMode = true
+        configureVoiceAudioSession()
         setState("listening", "Listening")
 
         mainHandler.post {
             try {
                 recognizer?.destroy()
-                recognizer = SpeechRecognizer.createSpeechRecognizer(context).apply {
-                    setRecognitionListener(recognitionListener)
-                }
+                recognizer = createRecognizer()
                 startListeningInternal(markListening = true)
                 startSilenceMonitor()
 
@@ -286,6 +356,13 @@ class TalkModePlugin : Plugin() {
                 })
             } catch (e: Exception) {
                 Log.e(TAG, "Failed to start", e)
+                // Recognizer creation failed AFTER the audio session was
+                // configured — release it so the earcon streams aren't left
+                // muted and the device isn't stuck in MODE_IN_COMMUNICATION.
+                enabled = false
+                listeningMode = false
+                releaseVoiceAudioSession()
+                setState("idle", "Off")
                 call.resolve(JSObject().apply {
                     put("started", false)
                     put("error", e.message ?: "Failed to start")
@@ -307,6 +384,10 @@ class TalkModePlugin : Plugin() {
         lastTranscript = ""
         lastHeardAtMs = null
 
+        // Release any raw-PCM capture; `enabled` is already false so this won't
+        // re-arm SpeechRecognizer.
+        stopAudioFramesInternal()
+
         mainHandler.post {
             recognizer?.cancel()
             recognizer?.destroy()
@@ -314,6 +395,7 @@ class TalkModePlugin : Plugin() {
         }
 
         stopSpeakingInternal()
+        releaseVoiceAudioSession()
         setState("idle", "Off")
         call.resolve()
     }
@@ -364,16 +446,18 @@ class TalkModePlugin : Plugin() {
         }
 
         val useSystemTts = call.getBoolean("useSystemTts", false) ?: false
+        val useLocalInferenceTts = call.getBoolean("useLocalInferenceTts", false) ?: false
         val directive = call.getObject("directive")
 
         speakingJob = scope.launch {
-            speakInternal(text, useSystemTts, directive, call)
+            speakInternal(text, useSystemTts, useLocalInferenceTts, directive, call)
         }
     }
 
     @PluginMethod
     fun stopSpeaking(call: PluginCall) {
         val interruptedAt = computeInterruptedAt()
+        lastInterruptedAtSeconds = interruptedAt
         stopSpeakingInternal()
         call.resolve(JSObject().apply {
             if (interruptedAt != null) {
@@ -408,6 +492,277 @@ class TalkModePlugin : Plugin() {
         call.resolve(buildPermissionResult())
     }
 
+    // ── Raw PCM frame capture (diarization / VAD / wake-word) ────────────
+
+    @PluginMethod
+    fun startAudioFrames(call: PluginCall) {
+        if (getPermissionState("microphone") != PermissionState.GRANTED) {
+            requestPermissionForAlias("microphone", call, "handleStartAudioFramesPermission")
+            return
+        }
+        startAudioFramesInternal(call)
+    }
+
+    @PermissionCallback
+    private fun handleStartAudioFramesPermission(call: PluginCall) {
+        if (getPermissionState("microphone") == PermissionState.GRANTED) {
+            startAudioFramesInternal(call)
+        } else {
+            call.resolve(JSObject().apply {
+                put("started", false)
+                put("error", "Microphone permission denied")
+            })
+        }
+    }
+
+    private fun startAudioFramesInternal(call: PluginCall) {
+        if (audioFrameRunning.get()) {
+            call.resolve(JSObject().apply {
+                put("started", true)
+                put("sampleRate", lastFrameSampleRate)
+                put("frameSamples", lastFrameSamples)
+                put("suspendedStt", sttSuspendedForFrames)
+            })
+            return
+        }
+
+        val requestedRate = call.getInt("sampleRate") ?: DEFAULT_FRAME_SAMPLE_RATE
+        val frameMs = call.getInt("frameMs") ?: DEFAULT_FRAME_MS
+        // SpeechRecognizer (SODA) holds the mic; a parallel AudioRecord on the
+        // same input fails on virtually every device. Suspend it for the
+        // duration of capture and remember to resume on stop.
+        val wasListening = isListening || listeningMode
+        if (wasListening) {
+            suspendSpeechRecognizerForFrames()
+        }
+
+        val record = try {
+            openAudioRecord(requestedRate)
+        } catch (e: Exception) {
+            Log.e(TAG, "AudioRecord open failed", e)
+            if (sttSuspendedForFrames) resumeSpeechRecognizerAfterFrames()
+            call.resolve(JSObject().apply {
+                put("started", false)
+                put("error", e.message ?: "AudioRecord open failed")
+            })
+            return
+        }
+
+        val actualRate = record.sampleRate
+        val frameSamples = max(1, actualRate * frameMs / 1000)
+        audioRecord = record
+        lastFrameSampleRate = actualRate
+        lastFrameSamples = frameSamples
+
+        try {
+            record.startRecording()
+        } catch (e: Exception) {
+            Log.e(TAG, "AudioRecord startRecording failed", e)
+            releaseAudioRecord()
+            if (sttSuspendedForFrames) resumeSpeechRecognizerAfterFrames()
+            call.resolve(JSObject().apply {
+                put("started", false)
+                put("error", e.message ?: "AudioRecord start failed")
+            })
+            return
+        }
+
+        if (record.recordingState != AudioRecord.RECORDSTATE_RECORDING) {
+            Log.e(TAG, "AudioRecord did not enter RECORDING state")
+            releaseAudioRecord()
+            if (sttSuspendedForFrames) resumeSpeechRecognizerAfterFrames()
+            call.resolve(JSObject().apply {
+                put("started", false)
+                put("error", "AudioRecord did not start (mic likely held by SpeechRecognizer)")
+            })
+            return
+        }
+
+        audioFrameRunning.set(true)
+        launchFrameLoop(record, frameSamples)
+
+        call.resolve(JSObject().apply {
+            put("started", true)
+            put("sampleRate", actualRate)
+            put("frameSamples", frameSamples)
+            put("suspendedStt", sttSuspendedForFrames)
+        })
+    }
+
+    @PluginMethod
+    fun stopAudioFrames(call: PluginCall) {
+        stopAudioFramesInternal()
+        call.resolve()
+    }
+
+    @PluginMethod
+    fun isCapturingAudioFrames(call: PluginCall) {
+        call.resolve(JSObject().apply {
+            put("capturing", audioFrameRunning.get())
+        })
+    }
+
+    /**
+     * Open a 16 kHz mono 16-bit AudioRecord. Tries VOICE_RECOGNITION first (the
+     * pre-processing-light source diarization wants), then falls back to MIC.
+     */
+    private fun openAudioRecord(sampleRate: Int): AudioRecord {
+        val minBuffer = AudioRecord.getMinBufferSize(
+            sampleRate,
+            AudioFormat.CHANNEL_IN_MONO,
+            AudioFormat.ENCODING_PCM_16BIT
+        )
+        if (minBuffer <= 0) {
+            throw IllegalStateException("AudioRecord min buffer invalid ($minBuffer) for ${sampleRate}Hz")
+        }
+        val bufferBytes = max(minBuffer * 2, 4 * 1024)
+        val sources = intArrayOf(
+            MediaRecorder.AudioSource.VOICE_RECOGNITION,
+            MediaRecorder.AudioSource.MIC,
+        )
+        var lastError: Throwable? = null
+        for (source in sources) {
+            try {
+                @Suppress("MissingPermission")
+                val record = AudioRecord(
+                    source,
+                    sampleRate,
+                    AudioFormat.CHANNEL_IN_MONO,
+                    AudioFormat.ENCODING_PCM_16BIT,
+                    bufferBytes
+                )
+                if (record.state == AudioRecord.STATE_INITIALIZED) {
+                    return record
+                }
+                record.release()
+                lastError = IllegalStateException("AudioRecord uninitialized for source $source")
+            } catch (e: Exception) {
+                lastError = e
+            }
+        }
+        throw IllegalStateException(
+            "AudioRecord could not initialize at ${sampleRate}Hz",
+            lastError
+        )
+    }
+
+    private fun launchFrameLoop(record: AudioRecord, frameSamples: Int) {
+        audioFrameJob?.cancel()
+        // IO dispatcher: a tight blocking read loop must not sit on the main
+        // thread. Frames are marshalled to JS via notifyListeners (thread-safe).
+        audioFrameJob = scope.launch(Dispatchers.IO) {
+            val buffer = ShortArray(frameSamples)
+            val bytes = ByteArray(frameSamples * 2)
+            var frameIndex = 0L
+            try {
+                while (audioFrameRunning.get() && isActive) {
+                    val read = record.read(buffer, 0, frameSamples)
+                    if (read <= 0) {
+                        // ERROR_INVALID_OPERATION (-3) / ERROR_BAD_VALUE (-2):
+                        // the record was released or the mic was taken; stop.
+                        if (read < 0) break
+                        continue
+                    }
+                    var sumSquares = 0.0
+                    var b = 0
+                    for (i in 0 until read) {
+                        val s = buffer[i].toInt()
+                        bytes[b] = (s and 0xff).toByte()
+                        bytes[b + 1] = ((s shr 8) and 0xff).toByte()
+                        b += 2
+                        sumSquares += (s.toDouble() * s.toDouble())
+                    }
+                    val rms = if (read > 0) {
+                        Math.sqrt(sumSquares / read) / 32768.0
+                    } else 0.0
+                    val pcmBase64 = Base64.encodeToString(
+                        bytes, 0, read * 2, Base64.NO_WRAP
+                    )
+                    val idx = frameIndex
+                    frameIndex += 1
+                    val ts = SystemClock.elapsedRealtime()
+                    notifyListeners("audioFrame", JSObject().apply {
+                        put("pcm16", pcmBase64)
+                        put("sampleRate", record.sampleRate)
+                        put("channels", 1)
+                        put("samples", read)
+                        put("rms", rms)
+                        put("timestamp", ts)
+                        put("frameIndex", idx)
+                    })
+                }
+            } catch (e: Throwable) {
+                Log.e(TAG, "Audio frame loop error", e)
+                notifyListeners("error", JSObject().apply {
+                    put("message", "Audio frame capture stopped: ${e.message}")
+                    put("fatal", false)
+                })
+            }
+        }
+    }
+
+    private fun stopAudioFramesInternal() {
+        if (!audioFrameRunning.getAndSet(false) && audioRecord == null) {
+            return
+        }
+        audioFrameJob?.cancel()
+        audioFrameJob = null
+        releaseAudioRecord()
+        if (sttSuspendedForFrames) {
+            resumeSpeechRecognizerAfterFrames()
+        }
+    }
+
+    private fun releaseAudioRecord() {
+        val record = audioRecord ?: return
+        audioRecord = null
+        try {
+            if (record.recordingState == AudioRecord.RECORDSTATE_RECORDING) {
+                record.stop()
+            }
+        } catch (_: Throwable) {
+        }
+        try {
+            record.release()
+        } catch (_: Throwable) {
+        }
+    }
+
+    /** Suspend SpeechRecognizer so AudioRecord can own the mic. */
+    private fun suspendSpeechRecognizerForFrames() {
+        sttSuspendedForFrames = true
+        listeningMode = false
+        isListening = false
+        restartJob?.cancel()
+        silenceJob?.cancel()
+        mainHandler.post {
+            try {
+                recognizer?.cancel()
+                recognizer?.destroy()
+            } catch (_: Throwable) {
+            }
+            recognizer = null
+        }
+    }
+
+    /** Re-arm SpeechRecognizer after frame capture ends, if a session is active. */
+    private fun resumeSpeechRecognizerAfterFrames() {
+        sttSuspendedForFrames = false
+        if (!enabled || stopRequested) return
+        listeningMode = true
+        mainHandler.post {
+            try {
+                if (!SpeechRecognizer.isRecognitionAvailable(context)) return@post
+                recognizer?.destroy()
+                recognizer = createRecognizer()
+                startListeningInternal(markListening = true)
+                startSilenceMonitor()
+            } catch (e: Exception) {
+                Log.e(TAG, "Failed to resume STT after frames", e)
+            }
+        }
+    }
+
     // ── Config ──────────────────────────────────────────────────────────
 
     private fun applyConfig(config: JSObject) {
@@ -462,6 +817,13 @@ class TalkModePlugin : Plugin() {
             putExtra(RecognizerIntent.EXTRA_PARTIAL_RESULTS, true)
             putExtra(RecognizerIntent.EXTRA_MAX_RESULTS, 3)
             putExtra(RecognizerIntent.EXTRA_CALLING_PACKAGE, context.packageName)
+            // On-device recognizer (no network round-trip; works offline). The
+            // platform recognizer's open/close cadence during continuous use is
+            // intrinsic and not controllable via the silence-length extras (the
+            // on-device SODA engine ignores them); we silence the AUDIBLE part of
+            // that churn by muting the earcon streams for the session instead
+            // (see configureVoiceAudioSession).
+            putExtra(RecognizerIntent.EXTRA_PREFER_OFFLINE, true)
             sttLanguage?.let { putExtra(RecognizerIntent.EXTRA_LANGUAGE, it) }
         }
 
@@ -477,6 +839,28 @@ class TalkModePlugin : Plugin() {
         }
     }
 
+    /**
+     * Create the speech recognizer. Prefer the API-31+ ON-DEVICE recognizer
+     * (in-process SODA): it plays NO start/error earcons, eliminating the
+     * audible "open"/"failure" beeps that came from the system
+     * com.google.android.tts recognizer service (which also can't be muted
+     * without ACCESS_NOTIFICATION_POLICY / STREAM_SYSTEM_ENFORCED control we
+     * don't hold). Falls back to the system recognizer when on-device SODA is
+     * unavailable.
+     */
+    private fun createRecognizer(): SpeechRecognizer {
+        val rec = if (
+            Build.VERSION.SDK_INT >= Build.VERSION_CODES.S &&
+            SpeechRecognizer.isOnDeviceRecognitionAvailable(context)
+        ) {
+            SpeechRecognizer.createOnDeviceSpeechRecognizer(context)
+        } else {
+            SpeechRecognizer.createSpeechRecognizer(context)
+        }
+        rec.setRecognitionListener(recognitionListener)
+        return rec
+    }
+
     private fun scheduleRestart(delayMs: Long = 350) {
         if (stopRequested) return
         restartJob?.cancel()
@@ -515,13 +899,14 @@ class TalkModePlugin : Plugin() {
         val elapsed = SystemClock.elapsedRealtime() - lastHeard
         if (elapsed < silenceWindowMs) return
 
-        // Finalize: emit a final transcript event
-        notifyListeners("transcript", JSObject().apply {
-            put("transcript", transcript)
-            put("isFinal", true)
-        })
+        // Finalize this turn (deduped against the recognizer's own onResults),
+        // then restart the recognizer so the next utterance is a CLEAN session —
+        // Android SpeechRecognizer accumulates within a session, so without the
+        // restart the next turn's partials would prepend the words we just sent.
         lastTranscript = ""
         lastHeardAtMs = null
+        emitFinalOnce(transcript)
+        scheduleRestart()
     }
 
     private fun handleTranscript(transcript: String, isFinal: Boolean) {
@@ -531,34 +916,71 @@ class TalkModePlugin : Plugin() {
         if (isSpeaking && interruptOnSpeech) {
             if (shouldInterrupt(transcript)) {
                 val interruptedAt = computeInterruptedAt()
-                stopSpeakingInternal()
                 lastInterruptedAtSeconds = interruptedAt
+                stopSpeakingInternal()
             }
             return
         }
 
         if (!isListening) return
 
-        if (transcript.isNotEmpty()) {
+        if (isFinal) {
+            // A real end-of-turn from the recognizer: emit once and clear the
+            // pending buffer so the silence monitor doesn't re-finalize the same
+            // words (the double-send bug).
+            lastTranscript = ""
+            lastHeardAtMs = null
+            emitFinalOnce(transcript)
+        } else {
             lastTranscript = transcript
             lastHeardAtMs = SystemClock.elapsedRealtime()
+            notifyListeners("transcript", JSObject().apply {
+                put("transcript", transcript)
+                put("isFinal", false)
+            })
         }
+    }
 
+    /**
+     * Emit a FINAL transcript exactly once. Both the recognizer's `onResults`
+     * and the silence monitor can finalize the same utterance; collapse them so
+     * the turn is sent a single time (a repeated final within 2s is dropped).
+     */
+    private fun emitFinalOnce(transcript: String) {
+        val text = transcript.trim()
+        if (text.isEmpty()) return
+        val now = SystemClock.elapsedRealtime()
+        if (text == lastEmittedFinal && now - lastEmittedFinalAtMs < 2000L) return
+        lastEmittedFinal = text
+        lastEmittedFinalAtMs = now
         notifyListeners("transcript", JSObject().apply {
-            put("transcript", transcript)
-            put("isFinal", isFinal)
+            put("transcript", text)
+            put("isFinal", true)
         })
     }
 
     /**
-     * Avoid false interrupts: don't interrupt if the heard text is just a
-     * substring of what we're currently speaking (echo from speaker).
+     * Decide whether heard speech should barge in on the agent's TTS. Tuned to
+     * avoid FALSE interrupts (which cut the reply mid-sentence and read as
+     * "intermittent audio"): a one-word ASR blip, background noise, or the
+     * agent's own voice bleeding back into the mic must NOT interrupt — only a
+     * genuine couple-of-words utterance from the user does.
      */
     private fun shouldInterrupt(transcript: String): Boolean {
         val trimmed = transcript.trim()
-        if (trimmed.length < 3) return false
-        val spoken = lastSpokenText?.lowercase()
-        if (spoken != null && spoken.contains(trimmed.lowercase())) return false
+        val lower = trimmed.lowercase()
+        val words = lower.split(Regex("\\s+")).filter { it.isNotBlank() }
+        // Need real intent: at least two words, or one long word (≥ 8 chars).
+        if (words.size < 2 && trimmed.length < 8) return false
+        val spoken = lastSpokenText?.lowercase() ?: return true
+        // Exact echo of what we're saying → speaker bleed, not the user.
+        if (spoken.contains(lower)) return false
+        // Fuzzy echo: if most of the heard words appear in the text we're
+        // currently speaking, treat it as echo (ASR mishears of our own audio).
+        val echoed = words.count { spoken.contains(it) }
+        if (words.isNotEmpty() && echoed.toDouble() / words.size >= 0.6) {
+            return false
+        }
         return true
     }
 
@@ -573,9 +995,7 @@ class TalkModePlugin : Plugin() {
             if (!SpeechRecognizer.isRecognitionAvailable(context)) return@post
             try {
                 if (recognizer == null) {
-                    recognizer = SpeechRecognizer.createSpeechRecognizer(context).apply {
-                        setRecognitionListener(recognitionListener)
-                    }
+                    recognizer = createRecognizer()
                 }
                 recognizer?.cancel()
                 startListeningInternal(markListening = false)
@@ -588,6 +1008,7 @@ class TalkModePlugin : Plugin() {
     private suspend fun speakInternal(
         text: String,
         forceSystemTts: Boolean,
+        useLocalInferenceTts: Boolean,
         directive: JSObject?,
         call: PluginCall
     ) {
@@ -596,6 +1017,7 @@ class TalkModePlugin : Plugin() {
         lastSpokenText = text
         speakStartTimeMs = SystemClock.elapsedRealtime()
         pcmStopRequested.set(false)
+        lastInterruptedAtSeconds = null
         setState("speaking", "Speaking")
 
         val effectiveVoiceId = directive.stringOrNull("voiceId")?.let(::resolveVoiceAlias) ?: voiceId
@@ -603,27 +1025,74 @@ class TalkModePlugin : Plugin() {
 
         notifyListeners("speaking", JSObject().apply {
             put("text", text)
-            put("isSystemTts", forceSystemTts || effectiveApiKey.isNullOrEmpty() || effectiveVoiceId.isNullOrEmpty())
+            put(
+                "isSystemTts",
+                !useLocalInferenceTts &&
+                    (forceSystemTts || effectiveApiKey.isNullOrEmpty() || effectiveVoiceId.isNullOrEmpty())
+            )
         })
 
         // Stop listening during speech (we keep recognizer for interrupt detection)
         mainHandler.post { recognizer?.stopListening() }
         ensureInterruptListener()
 
-        // Request audio focus
-        requestAudioFocus()
+        // Ensure the communication-mode session + audio focus are active even
+        // for a standalone speak() that wasn't preceded by start().
+        configureVoiceAudioSession()
+        // Re-assert loudspeaker routing right before playback. configureVoice…
+        // only routes on the FIRST activation; if the session was already up (the
+        // STT path opened it) the speaker route may have drifted, leaving TTS on
+        // the earpiece. Re-route here so replies are audible out the speaker.
+        audioManager?.let { routeVoiceOutput(it) }
 
         try {
-            val canUseElevenLabs = !forceSystemTts &&
+            val canUseLocalInference = useLocalInferenceTts && !forceSystemTts
+            val canUseElevenLabs = !canUseLocalInference &&
+                !forceSystemTts &&
                 !effectiveApiKey.isNullOrEmpty() &&
                 !effectiveVoiceId.isNullOrEmpty()
 
-            if (canUseElevenLabs) {
+            if (canUseLocalInference) {
+                try {
+                    streamAndPlayLocalInferenceTts(text, directive)
+
+                    if (!pcmStopRequested.get()) {
+                        call.resolve(JSObject().apply {
+                            put("completed", true)
+                            put("interrupted", false)
+                            put("usedSystemTts", false)
+                        })
+                    } else {
+                        call.resolve(JSObject().apply {
+                            put("completed", false)
+                            put("interrupted", true)
+                            put("usedSystemTts", false)
+                            lastInterruptedAtSeconds?.let { put("interruptedAt", it) }
+                        })
+                    }
+                } catch (e: Exception) {
+                    if (pcmStopRequested.get()) {
+                        call.resolve(JSObject().apply {
+                            put("completed", false)
+                            put("interrupted", true)
+                            put("usedSystemTts", false)
+                        })
+                    } else {
+                        // The on-device OmniVoice TTS assets aren't always staged
+                        // (it 502s "TEXT_TO_SPEECH not available"). Rather than go
+                        // silent — the JS browser-SpeechSynthesis fallback doesn't
+                        // exist in the Android WebView — fall back to the platform
+                        // TextToSpeech so replies are always spoken aloud.
+                        Log.w(TAG, "Local inference TTS failed, falling back to system TTS", e)
+                        speakWithSystemTts(text, call)
+                    }
+                }
+            } else if (canUseElevenLabs) {
                 try {
                     val request = buildElevenLabsRequest(text, directive)
                     streamAndPlayPcm(
-                        voiceId = effectiveVoiceId!!,
-                        apiKey = effectiveApiKey!!,
+                        voiceId = effectiveVoiceId,
+                        apiKey = effectiveApiKey,
                         request = request
                     )
 
@@ -665,13 +1134,16 @@ class TalkModePlugin : Plugin() {
                 put("error", e.message ?: "Speak failed")
             })
         } finally {
+            val wasInterrupted = pcmStopRequested.get()
+            val interruptedAt = lastInterruptedAtSeconds
             isSpeaking = false
             pcmStopRequested.set(false)
-            abandonAudioFocus()
 
             notifyListeners("speakComplete", JSObject().apply {
-                put("completed", !pcmStopRequested.get())
-                lastInterruptedAtSeconds?.let { put("interruptedAt", it) }
+                put("completed", !wasInterrupted)
+                if (wasInterrupted) {
+                    interruptedAt?.let { put("interruptedAt", it) }
+                }
             })
 
             if (enabled) {
@@ -679,6 +1151,8 @@ class TalkModePlugin : Plugin() {
                 setState("listening", "Listening")
                 mainHandler.post { startListeningInternal(markListening = true) }
             } else {
+                // Standalone speak (no active conversation): release the session.
+                releaseVoiceAudioSession()
                 setState("idle", "Off")
             }
         }
@@ -753,6 +1227,363 @@ class TalkModePlugin : Plugin() {
         return if (value == null || value === JSONObject.NULL) null else value.toString()
     }
 
+    private data class PcmStreamFormat(
+        val sampleRate: Int,
+        val channels: Int,
+        val bitsPerSample: Int,
+        val dataBytes: Int
+    )
+
+    /**
+     * Stream local-inference TTS from the embedded agent and play it natively.
+     *
+     * The agent currently returns a buffered WAV, but keeping playback in
+     * AudioTrack means this path is ready for a chunked PCM/WAV response without
+     * going back through WebView decodeAudioData.
+     */
+    private suspend fun streamAndPlayLocalInferenceTts(
+        text: String,
+        directive: JSObject?
+    ) = withContext(Dispatchers.IO) {
+        pcmStopRequested.set(false)
+        // Prefer the in-process fused Kokoro voice via the bionic inference host.
+        // Only if that host is unreachable (e.g. desktop/Electrobun, or a build
+        // without it) do we fall through to the HTTP agent endpoint.
+        if (streamAndPlayBionicKokoroTts(text, directive)) {
+            return@withContext
+        }
+        val conn = openLocalInferenceTtsConnection()
+        activePcmConnection = conn
+        try {
+            val payload = buildLocalInferenceTtsPayload(text, directive)
+            conn.outputStream.use { it.write(payload.toByteArray(Charsets.UTF_8)) }
+
+            val code = conn.responseCode
+            if (code >= 400) {
+                val errBody = conn.errorStream?.readBytes()?.toString(Charsets.UTF_8) ?: ""
+                throw IllegalStateException("Local inference TTS error: $code $errBody")
+            }
+
+            BufferedInputStream(conn.inputStream).use { input ->
+                val format = readWavPcmFormat(input)
+                val track = createPcmAudioTrack(format)
+                pcmTrack = track
+                track.play()
+
+                Log.d(
+                    TAG,
+                    "Local inference PCM play start sampleRate=${format.sampleRate} channels=${format.channels}"
+                )
+                notifyListeners("playbackStart", JSObject().apply {
+                    put("provider", "local-inference")
+                    put("sampleRate", format.sampleRate)
+                    put("channels", format.channels)
+                })
+                val framesWritten = writePcmStreamToTrack(input, track, format)
+                drainPcmTrack(track, framesWritten, format.sampleRate)
+                if (!pcmStopRequested.get()) {
+                    track.stop()
+                }
+                Log.d(TAG, "Local inference PCM play done frames=$framesWritten")
+            }
+        } finally {
+            cleanupPcmTrack()
+            if (activePcmConnection === conn) {
+                activePcmConnection = null
+            }
+            conn.disconnect()
+        }
+    }
+
+    /**
+     * Synthesize + play with the fused Kokoro-82M head in the bionic inference
+     * host (ElizaBionicInferenceServer, op "tts") over its abstract-namespace
+     * UDS. The host loads the same libelizainference that runs GPU text and
+     * synthesizes Kokoro PCM in-process — no musl agent, no HTTP, no 502 → no
+     * fallback to the platform TextToSpeech (the bug this fixes: the app was
+     * speaking with the Android system voice). Returns true on success; false if
+     * the host is unreachable so the caller can fall through.
+     */
+    private suspend fun streamAndPlayBionicKokoroTts(
+        text: String,
+        directive: JSObject?
+    ): Boolean = withContext(Dispatchers.IO) {
+        val trimmed = text.trim()
+        if (trimmed.isEmpty()) return@withContext false
+        val speed = (directive?.optDouble("speed", 1.0) ?: 1.0).toFloat()
+        val sock = LocalSocket()
+        try {
+            sock.connect(
+                LocalSocketAddress(BIONIC_INFER_SOCKET, LocalSocketAddress.Namespace.ABSTRACT)
+            )
+        } catch (e: Exception) {
+            Log.d(TAG, "bionic Kokoro TTS host unreachable: ${e.message}")
+            try { sock.close() } catch (_: Exception) {}
+            return@withContext false
+        }
+        try {
+            val req = JSONObject().apply {
+                put("op", "tts")
+                put("text", trimmed)
+                put("speed", speed.toDouble())
+            }.toString().toByteArray(Charsets.UTF_8)
+            DataOutputStream(sock.outputStream).apply {
+                writeInt(req.size) // big-endian length prefix
+                write(req)
+                flush()
+            }
+            val din = DataInputStream(sock.inputStream)
+            val len = din.readInt()
+            if (len <= 0 || len > 64 * 1024 * 1024) {
+                throw IllegalStateException("bionic TTS bad frame length $len")
+            }
+            val respBytes = ByteArray(len)
+            din.readFully(respBytes)
+            val resp = JSONObject(String(respBytes, Charsets.UTF_8))
+            if (!resp.optBoolean("ok", false)) {
+                throw IllegalStateException("bionic TTS error: ${resp.optString("error")}")
+            }
+            val sampleRate = resp.optInt("sampleRate", 24000)
+            val pcmF32 = Base64.decode(resp.getString("pcmBase64"), Base64.NO_WRAP)
+            // fp32 LE → int16 PCM (the play path is ENCODING_PCM_16BIT).
+            val fb = ByteBuffer.wrap(pcmF32).order(ByteOrder.LITTLE_ENDIAN).asFloatBuffer()
+            val nSamples = fb.remaining()
+            if (nSamples == 0) {
+                throw IllegalStateException("bionic TTS returned 0 samples")
+            }
+            val pcm16 = ByteArray(nSamples * 2)
+            val ob = ByteBuffer.wrap(pcm16).order(ByteOrder.LITTLE_ENDIAN)
+            for (i in 0 until nSamples) {
+                val s = (fb.get(i) * 32767f).coerceIn(-32768f, 32767f).toInt().toShort()
+                ob.putShort(s)
+            }
+            val format = PcmStreamFormat(sampleRate, 1, 16, pcm16.size)
+            val track = createPcmAudioTrack(format)
+            pcmTrack = track
+            track.play()
+            notifyListeners("playbackStart", JSObject().apply {
+                put("provider", "local-inference")
+                put("sampleRate", sampleRate)
+                put("channels", 1)
+            })
+            val framesWritten = writePcmStreamToTrack(
+                BufferedInputStream(ByteArrayInputStream(pcm16)), track, format
+            )
+            drainPcmTrack(track, framesWritten, sampleRate)
+            if (!pcmStopRequested.get()) track.stop()
+            Log.d(TAG, "bionic Kokoro TTS played $nSamples samples @ $sampleRate Hz")
+            true
+        } finally {
+            cleanupPcmTrack()
+            try { sock.close() } catch (_: Exception) {}
+        }
+    }
+
+    private fun openLocalInferenceTtsConnection(): HttpURLConnection {
+        val tokenFile = File(context.filesDir, "auth/local-agent-token")
+        val token = tokenFile.takeIf { it.isFile }?.readText()?.trim().orEmpty()
+        if (token.isEmpty()) {
+            throw IllegalStateException("Local agent auth token is missing")
+        }
+
+        val conn = URL(LOCAL_INFERENCE_TTS_URL).openConnection() as HttpURLConnection
+        conn.requestMethod = "POST"
+        conn.connectTimeout = 30_000
+        conn.readTimeout = 180_000
+        conn.setRequestProperty("Authorization", "Bearer $token")
+        conn.setRequestProperty("Content-Type", "application/json")
+        conn.setRequestProperty("Accept", "audio/wav")
+        conn.doOutput = true
+        return conn
+    }
+
+    private fun buildLocalInferenceTtsPayload(text: String, directive: JSObject?): String {
+        val payload = JSONObject()
+        payload.put("text", text)
+        directive.stringOrNull("voiceId")?.let { payload.put("voiceId", it) }
+        directive.stringOrNull("voice")?.let { payload.put("voice", it) }
+        directive.stringOrNull("modelId")?.let { payload.put("modelId", it) }
+        directive.stringOrNull("model")?.let { payload.put("model", it) }
+        val speed = directive?.optDouble("speed", Double.NaN)
+        if (speed != null && speed.isFinite() && speed > 0.0) {
+            payload.put("speed", speed)
+        }
+        return payload.toString()
+    }
+
+    private fun readExactly(input: BufferedInputStream, size: Int): ByteArray {
+        val bytes = ByteArray(size)
+        var offset = 0
+        while (offset < size) {
+            val read = input.read(bytes, offset, size - offset)
+            if (read < 0) {
+                throw IllegalStateException("Unexpected end of WAV stream")
+            }
+            offset += read
+        }
+        return bytes
+    }
+
+    private fun skipFully(input: BufferedInputStream, count: Int) {
+        var remaining = count
+        while (remaining > 0) {
+            val skipped = input.skip(remaining.toLong()).toInt()
+            if (skipped > 0) {
+                remaining -= skipped
+                continue
+            }
+            if (input.read() < 0) {
+                throw IllegalStateException("Unexpected end of WAV stream")
+            }
+            remaining -= 1
+        }
+    }
+
+    private fun littleEndianShort(bytes: ByteArray, offset: Int): Int {
+        return (bytes[offset].toInt() and 0xff) or
+            ((bytes[offset + 1].toInt() and 0xff) shl 8)
+    }
+
+    private fun littleEndianInt(bytes: ByteArray, offset: Int): Int {
+        return (bytes[offset].toInt() and 0xff) or
+            ((bytes[offset + 1].toInt() and 0xff) shl 8) or
+            ((bytes[offset + 2].toInt() and 0xff) shl 16) or
+            ((bytes[offset + 3].toInt() and 0xff) shl 24)
+    }
+
+    private fun chunkId(bytes: ByteArray): String {
+        return String(bytes, 0, 4, Charsets.US_ASCII)
+    }
+
+    private fun readWavPcmFormat(input: BufferedInputStream): PcmStreamFormat {
+        val riff = readExactly(input, 12)
+        if (
+            String(riff, 0, 4, Charsets.US_ASCII) != "RIFF" ||
+            String(riff, 8, 4, Charsets.US_ASCII) != "WAVE"
+        ) {
+            throw IllegalStateException("Local inference TTS returned non-WAV audio")
+        }
+
+        var format: PcmStreamFormat? = null
+        while (true) {
+            val header = readExactly(input, 8)
+            val id = chunkId(header)
+            val size = littleEndianInt(header, 4)
+            if (size < 0) {
+                throw IllegalStateException("Invalid WAV chunk size for $id")
+            }
+
+            if (id == "fmt ") {
+                val fmt = readExactly(input, size)
+                if (fmt.size < 16) {
+                    throw IllegalStateException("Invalid WAV fmt chunk")
+                }
+                val audioFormat = littleEndianShort(fmt, 0)
+                val channels = littleEndianShort(fmt, 2)
+                val sampleRate = littleEndianInt(fmt, 4)
+                val bitsPerSample = littleEndianShort(fmt, 14)
+                if (audioFormat != 1) {
+                    throw IllegalStateException("Only PCM WAV is supported, got format=$audioFormat")
+                }
+                if (bitsPerSample != 16) {
+                    throw IllegalStateException("Only 16-bit PCM WAV is supported, got bits=$bitsPerSample")
+                }
+                if (channels !in 1..2 || sampleRate <= 0) {
+                    throw IllegalStateException("Invalid WAV format sampleRate=$sampleRate channels=$channels")
+                }
+                format = PcmStreamFormat(sampleRate, channels, bitsPerSample, 0)
+                if (size % 2 == 1) skipFully(input, 1)
+                continue
+            }
+
+            if (id == "data") {
+                val parsed = format ?: throw IllegalStateException("WAV data arrived before fmt chunk")
+                return parsed.copy(dataBytes = size)
+            }
+
+            skipFully(input, size)
+            if (size % 2 == 1) skipFully(input, 1)
+        }
+    }
+
+    private fun createPcmAudioTrack(format: PcmStreamFormat): AudioTrack {
+        val channelMask = when (format.channels) {
+            1 -> AudioFormat.CHANNEL_OUT_MONO
+            2 -> AudioFormat.CHANNEL_OUT_STEREO
+            else -> throw IllegalStateException("Unsupported PCM channel count ${format.channels}")
+        }
+        val minBuffer = AudioTrack.getMinBufferSize(
+            format.sampleRate,
+            channelMask,
+            AudioFormat.ENCODING_PCM_16BIT
+        )
+        if (minBuffer <= 0) {
+            throw IllegalStateException("AudioTrack buffer size invalid: $minBuffer")
+        }
+        val bufferSize = max(minBuffer * 2, 8 * 1024)
+        val track = AudioTrack.Builder()
+            .setAudioAttributes(voiceAudioAttributes())
+            .setAudioFormat(
+                AudioFormat.Builder()
+                    .setEncoding(AudioFormat.ENCODING_PCM_16BIT)
+                    .setSampleRate(format.sampleRate)
+                    .setChannelMask(channelMask)
+                    .build()
+            )
+            .setBufferSizeInBytes(bufferSize)
+            .setTransferMode(AudioTrack.MODE_STREAM)
+            .build()
+
+        if (track.state != AudioTrack.STATE_INITIALIZED) {
+            track.release()
+            throw IllegalStateException("AudioTrack init failed")
+        }
+        return track
+    }
+
+    private fun writePcmStreamToTrack(
+        input: BufferedInputStream,
+        track: AudioTrack,
+        format: PcmStreamFormat
+    ): Long {
+        val bytesPerFrame = format.channels * (format.bitsPerSample / 8)
+        var bytesWrittenTotal = 0L
+        var remainingBytes = format.dataBytes
+        val buffer = ByteArray(8 * 1024)
+        while (remainingBytes > 0) {
+            if (pcmStopRequested.get()) break
+            val requestBytes = if (remainingBytes < buffer.size) remainingBytes else buffer.size
+            val bytesRead = input.read(buffer, 0, requestBytes)
+            if (bytesRead <= 0) break
+            remainingBytes -= bytesRead
+
+            var offset = 0
+            while (offset < bytesRead) {
+                if (pcmStopRequested.get()) break
+                val wrote = track.write(buffer, offset, bytesRead - offset)
+                if (wrote <= 0) {
+                    throw IllegalStateException("AudioTrack write failed: $wrote")
+                }
+                offset += wrote
+                bytesWrittenTotal += wrote.toLong()
+            }
+        }
+        return if (bytesPerFrame > 0) bytesWrittenTotal / bytesPerFrame else 0L
+    }
+
+    private fun drainPcmTrack(track: AudioTrack, framesWritten: Long, sampleRate: Int) {
+        if (framesWritten <= 0L || sampleRate <= 0) return
+        val maxDrainMs = (framesWritten * 1000L / sampleRate).coerceAtMost(30_000L) + 1_000L
+        val deadline = SystemClock.elapsedRealtime() + maxDrainMs
+        while (
+            !pcmStopRequested.get() &&
+            track.playbackHeadPosition.toLong() < framesWritten &&
+            SystemClock.elapsedRealtime() < deadline
+        ) {
+            SystemClock.sleep(20)
+        }
+    }
+
     /**
      * Stream PCM audio from ElevenLabs and play via AudioTrack.
      * Ported from classic TalkModeManager with proper offset-based writes.
@@ -776,12 +1607,7 @@ class TalkModePlugin : Plugin() {
 
         val bufferSize = max(minBuffer * 2, 8 * 1024)
         val track = AudioTrack.Builder()
-            .setAudioAttributes(
-                AudioAttributes.Builder()
-                    .setUsage(AudioAttributes.USAGE_ASSISTANT)
-                    .setContentType(AudioAttributes.CONTENT_TYPE_SPEECH)
-                    .build()
-            )
+            .setAudioAttributes(voiceAudioAttributes())
             .setAudioFormat(
                 AudioFormat.Builder()
                     .setEncoding(AudioFormat.ENCODING_PCM_16BIT)
@@ -802,6 +1628,7 @@ class TalkModePlugin : Plugin() {
 
         Log.d(TAG, "PCM play start sampleRate=$sampleRate bufferSize=$bufferSize")
         val conn = openTtsConnection(voiceId, apiKey, request)
+        activePcmConnection = conn
         try {
             val payload = buildRequestPayload(request)
             conn.outputStream.use { it.write(payload.toByteArray()) }
@@ -845,6 +1672,9 @@ class TalkModePlugin : Plugin() {
             Log.d(TAG, "PCM play done")
         } finally {
             cleanupPcmTrack()
+            if (activePcmConnection === conn) {
+                activePcmConnection = null
+            }
             conn.disconnect()
         }
     }
@@ -970,43 +1800,125 @@ class TalkModePlugin : Plugin() {
         }
     }
 
-    // ── Audio focus ─────────────────────────────────────────────────────
+    // ── Voice audio session ─────────────────────────────────────────────
+    //
+    // The Android analog of the iOS `.playAndRecord` / `.voiceChat` /
+    // `.defaultToSpeaker` session. Putting the device in MODE_IN_COMMUNICATION
+    // for the whole conversation routes capture + playback through the
+    // telephony path, which engages the platform hardware AEC so TTS coming out
+    // the speaker is cancelled from the mic (the core fix for the mic+speaker
+    // echo loop in hands-free mode). We also hold voice-communication audio
+    // focus and route to the loudspeaker (unless a headset is connected) so
+    // hands-free playback is audible.
+
+    private fun voiceAudioAttributes(): AudioAttributes =
+        AudioAttributes.Builder()
+            .setUsage(AudioAttributes.USAGE_VOICE_COMMUNICATION)
+            .setContentType(AudioAttributes.CONTENT_TYPE_SPEECH)
+            .build()
 
-    private fun requestAudioFocus() {
+    private fun configureVoiceAudioSession() {
+        if (audioSessionActive) return
         val am = audioManager ?: return
-        val focusListener = AudioManager.OnAudioFocusChangeListener { focusChange ->
-            when (focusChange) {
-                AudioManager.AUDIOFOCUS_LOSS,
-                AudioManager.AUDIOFOCUS_LOSS_TRANSIENT -> {
-                    // Another app took audio; stop speaking if we are
-                    if (isSpeaking) {
-                        stopSpeakingInternal()
-                    }
+
+        savedAudioMode = am.mode
+        @Suppress("DEPRECATION")
+        savedSpeakerphoneOn = am.isSpeakerphoneOn
+
+        val request = AudioFocusRequest.Builder(AudioManager.AUDIOFOCUS_GAIN_TRANSIENT_EXCLUSIVE)
+            .setAudioAttributes(voiceAudioAttributes())
+            .setOnAudioFocusChangeListener { focusChange ->
+                if (
+                    focusChange == AudioManager.AUDIOFOCUS_LOSS ||
+                    focusChange == AudioManager.AUDIOFOCUS_LOSS_TRANSIENT
+                ) {
+                    // Another app took audio; stop speaking if we are.
+                    if (isSpeaking) stopSpeakingInternal()
                 }
             }
+            .build()
+        audioFocusRequest = request
+        am.requestAudioFocus(request)
+
+        am.mode = AudioManager.MODE_IN_COMMUNICATION
+        routeVoiceOutput(am)
+        muteEarconStreams(am)
+        audioSessionActive = true
+        Log.d(TAG, "Voice audio session active (communication mode)")
+    }
+
+    /** Mute the recognizer earcon streams for the session; idempotent. */
+    private fun muteEarconStreams(am: AudioManager) {
+        if (earconStreamsMuted) return
+        for (stream in earconStreams) {
+            try {
+                am.adjustStreamVolume(stream, AudioManager.ADJUST_MUTE, 0)
+            } catch (_: Throwable) {
+                // Some OEMs disallow muting certain streams without DND access.
+            }
+        }
+        earconStreamsMuted = true
+    }
+
+    private fun unmuteEarconStreams(am: AudioManager) {
+        if (!earconStreamsMuted) return
+        for (stream in earconStreams) {
+            try {
+                am.adjustStreamVolume(stream, AudioManager.ADJUST_UNMUTE, 0)
+            } catch (_: Throwable) {}
         }
-        audioFocusRequest = focusListener
+        earconStreamsMuted = false
+    }
 
+    /**
+     * Default playback to the loudspeaker for hands-free use, but let a wired or
+     * Bluetooth headset win — the iOS `.defaultToSpeaker` semantic.
+     */
+    private fun routeVoiceOutput(am: AudioManager) {
+        val hasHeadset = am.getDevices(AudioManager.GET_DEVICES_OUTPUTS).any { device ->
+            device.type == AudioDeviceInfo.TYPE_WIRED_HEADSET ||
+                device.type == AudioDeviceInfo.TYPE_WIRED_HEADPHONES ||
+                device.type == AudioDeviceInfo.TYPE_USB_HEADSET ||
+                device.type == AudioDeviceInfo.TYPE_BLUETOOTH_SCO ||
+                device.type == AudioDeviceInfo.TYPE_BLUETOOTH_A2DP
+        }
+        if (hasHeadset) {
+            if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.S) am.clearCommunicationDevice()
+            @Suppress("DEPRECATION")
+            am.isSpeakerphoneOn = false
+            return
+        }
+        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.S) {
+            val speaker = am.availableCommunicationDevices.firstOrNull {
+                it.type == AudioDeviceInfo.TYPE_BUILTIN_SPEAKER
+            }
+            if (speaker != null && am.setCommunicationDevice(speaker)) return
+        }
         @Suppress("DEPRECATION")
-        am.requestAudioFocus(
-            focusListener,
-            AudioManager.STREAM_MUSIC,
-            AudioManager.AUDIOFOCUS_GAIN_TRANSIENT_MAY_DUCK
-        )
+        am.isSpeakerphoneOn = true
     }
 
-    private fun abandonAudioFocus() {
+    private fun releaseVoiceAudioSession() {
+        if (!audioSessionActive) return
         val am = audioManager ?: return
-        val listener = audioFocusRequest ?: return
-        @Suppress("DEPRECATION")
-        am.abandonAudioFocus(listener)
+        unmuteEarconStreams(am)
+        audioFocusRequest?.let { am.abandonAudioFocusRequest(it) }
         audioFocusRequest = null
+        if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.S) am.clearCommunicationDevice()
+        @Suppress("DEPRECATION")
+        am.isSpeakerphoneOn = savedSpeakerphoneOn
+        am.mode = savedAudioMode
+        audioSessionActive = false
+        Log.d(TAG, "Voice audio session released")
     }
 
     // ── Cleanup helpers ─────────────────────────────────────────────────
 
     private fun stopSpeakingInternal() {
         pcmStopRequested.set(true)
+        val conn = activePcmConnection
+        activePcmConnection = null
+        conn?.disconnect()
         cleanupPcmTrack()
         systemTts?.stop()
         systemTtsPending?.cancel()
@@ -1162,6 +2074,9 @@ class TalkModePlugin : Plugin() {
     }
 
     private fun isPermissionGranted(permission: String): Boolean {
+        if (permission == Manifest.permission.RECORD_AUDIO) {
+            return context.checkSelfPermission(permission) == PackageManager.PERMISSION_GRANTED
+        }
         return getPermissionState(permission) == com.getcapacitor.PermissionState.GRANTED
     }
 
@@ -1176,10 +2091,13 @@ class TalkModePlugin : Plugin() {
         systemTts?.shutdown()
         systemTts = null
         cleanupPcmTrack()
+        audioFrameRunning.set(false)
+        audioFrameJob?.cancel()
+        releaseAudioRecord()
         silenceJob?.cancel()
         restartJob?.cancel()
         speakingJob?.cancel()
-        abandonAudioFocus()
+        releaseVoiceAudioSession()
         scope.cancel()
     }