feat(voice): replace amplitude silence detection with Silero VAD

Use VadHandler from the vad package (Silero VAD v5 ONNX) for speech
detection instead of amplitude thresholds. VAD runs its own PCM
stream while the file recorder captures AAC-LC for Whisper. Grace
period starts on speech-start, auto-stop on speech-end after grace.
No-speech guard shows error on manual stop without detected speech.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

lib/providers/voice_provider.dart

@@ -8,6 +8,7 @@ import 'package:just_audio/just_audio.dart';
 import 'package:path_provider/path_provider.dart';
 import 'package:permission_handler/permission_handler.dart';
 import 'package:record/record.dart';
+import 'package:vad/vad.dart';
 
 import 'api_client_provider.dart';
 
@@ -98,25 +99,13 @@ class VoiceNotifier extends Notifier<VoiceState> {
   AudioPlayer? _player;
   StreamSubscription<Amplitude>? _amplitudeSubscription;
 
-  // Recording / silence detection
-  //
-  // Silence threshold is dynamic: we track the session peak dBFS and treat
-  // "silent" as "current level is at least _dropFromPeakDb below peak."
-  // This auto-calibrates to whatever mic + room the user is on rather than
-  // assuming a fixed ambient level. Until the peak climbs above
-  // _dynamicArmDb we fall back to a conservative static threshold so a
-  // quiet room doesn't spin forever. A grace period at the start of the
-  // recording gives the user time to begin speaking before silence checks
-  // arm.
-  int _recordingStartMs = 0;
-  int _silenceMs = 0;
-  double _peakDb = -100.0;
-  static const _fallbackThresholdDb = -35.0;
-  static const _dropFromPeakDb = 15.0;
-  static const _dynamicArmDb = -20.0;
-  static const _graceMs = 1500;
-  static const _silenceDurationMs = 2000;
-  static const _minRecordingMs = 300;
+  // VAD-based speech detection
+  VadHandler? _vadHandler;
+  StreamSubscription<void>? _vadSpeechStartSub;
+  StreamSubscription<List<double>>? _vadSpeechEndSub;
+  bool _speechDetected = false;
+  int _speechStartMs = 0;
+  static const _vadGraceMs = 1500;
 
   // Voice mode callbacks
   Future<void> Function(String transcript)? _onTranscript;
@@ -164,6 +153,9 @@ class VoiceNotifier extends Notifier<VoiceState> {
     required void Function(String message) onError,
   }) async {
     if (state.voiceModeActive) {
+      if (state.mode == VoiceMode.recording && !_speechDetected) {
+        onError('No speech detected');
+      }
       exitVoiceMode();
       return;
     }
@@ -211,12 +203,14 @@ class VoiceNotifier extends Notifier<VoiceState> {
     _amplitudeSubscription?.cancel();
     _amplitudeSubscription = null;
     _recorder?.stop();
+    _stopVad();
     _player?.stop();
     _ttsQueue.clear();
     _ttsPlaying = false;
     _sentenceBuffer = '';
     _lastSeenLength = 0;
     _streamComplete = false;
+    _speechDetected = false;
     _onTranscript = null;
     _onError = null;
     state = const VoiceState();
@@ -248,9 +242,8 @@ class VoiceNotifier extends Notifier<VoiceState> {
   Future<void> _startListening() async {
     if (!state.voiceModeActive) return;
 
-    _silenceMs = 0;
-    _peakDb = -100.0;
-    _recordingStartMs = DateTime.now().millisecondsSinceEpoch;
+    _speechDetected = false;
+    _speechStartMs = 0;
     state = state.copyWith(mode: VoiceMode.recording);
 
     final dir = _tempDir ?? await getTemporaryDirectory();
@@ -258,9 +251,6 @@ class VoiceNotifier extends Notifier<VoiceState> {
         '${dir.path}/voice_rec_${DateTime.now().millisecondsSinceEpoch}.m4a';
 
     try {
-      // Recreate the recorder each session — the record package can leave
-      // the native AudioRecord in a bad state after stop/error cycles,
-      // and a stale instance is the most common cause of "could not start".
       _recorder?.dispose();
       _recorder = AudioRecorder();
 
@@ -279,53 +269,53 @@ class VoiceNotifier extends Notifier<VoiceState> {
       _amplitudeSubscription = _recorder!
           .onAmplitudeChanged(const Duration(milliseconds: 200))
           .listen(_onAmplitude);
+
+      // VAD for speech detection — uses its own AudioRecorder internally
+      await _stopVad();
+      _vadHandler = VadHandler.create();
+      _vadSpeechStartSub = _vadHandler!.onSpeechStart.listen((_) {
+        if (!_speechDetected) {
+          _speechDetected = true;
+          _speechStartMs = DateTime.now().millisecondsSinceEpoch;
+        }
+      });
+      _vadSpeechEndSub = _vadHandler!.onSpeechEnd.listen((_) {
+        if (!state.voiceModeActive) return;
+        final now = DateTime.now().millisecondsSinceEpoch;
+        final sinceStart = _speechStartMs > 0 ? now - _speechStartMs : 0;
+        if (_speechDetected && sinceStart >= _vadGraceMs) {
+          _amplitudeSubscription?.cancel();
+          _amplitudeSubscription = null;
+          _stopVad();
+          _handleSilence();
+        }
+      });
+      await _vadHandler!.startListening(model: 'v5');
     } catch (e) {
       _onError?.call('Microphone error: $e');
       exitVoiceMode();
     }
   }
 
+  Future<void> _stopVad() async {
+    await _vadSpeechStartSub?.cancel();
+    _vadSpeechStartSub = null;
+    await _vadSpeechEndSub?.cancel();
+    _vadSpeechEndSub = null;
+    if (_vadHandler != null) {
+      await _vadHandler!.dispose();
+      _vadHandler = null;
+    }
+  }
+
   void _onAmplitude(Amplitude event) {
     if (!state.voiceModeActive) return;
-
     final db = event.current;
-    final validDb = !(db.isNaN || db.isInfinite);
-
-    // Normalize dB to 0..1 for the pulse animation. -60dB is dead-quiet,
-    // 0dB is peak; we clamp and bias so the button visibly breathes even
-    // on soft speech without exploding on loud input.
-    if (validDb) {
+    if (db.isNaN || db.isInfinite) return;
     final norm = ((db + 60.0) / 60.0).clamp(0.0, 1.0);
     if ((norm - state.amplitude).abs() > 0.02) {
       state = state.copyWith(amplitude: norm);
     }
-      if (db > _peakDb) _peakDb = db;
-    }
-
-    final elapsed =
-        DateTime.now().millisecondsSinceEpoch - _recordingStartMs;
-    // Grace period at the start — let the user begin speaking before we
-    // start silence-counting. Also suppresses any false triggers while
-    // the native recorder is still warming up.
-    if (elapsed < _graceMs) return;
-    if (elapsed < _minRecordingMs) return;
-
-    // Dynamic threshold once we've seen real speech; static fallback
-    // before that so a dead-silent session doesn't spin forever.
-    final threshold =
-        _peakDb > _dynamicArmDb ? _peakDb - _dropFromPeakDb : _fallbackThresholdDb;
-    final isSilent = !validDb || db < threshold;
-
-    if (isSilent) {
-      _silenceMs += 200;
-      if (_silenceMs >= _silenceDurationMs) {
-        _amplitudeSubscription?.cancel();
-        _amplitudeSubscription = null;
-        _handleSilence();
-      }
-    } else {
-      _silenceMs = 0;
-    }
   }
 
   Future<void> _handleSilence() async {