Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.runanywhere.ai/llms.txt

Use this file to discover all available pages before exploring further.

Voice Activity Detection (VAD) automatically detects speech in audio streams, enabling hands-free voice interfaces.

​
Overview

VAD determines when a user is speaking vs. silent, which is essential for:
  • Knowing when to start/stop recording
  • Triggering transcription at the right time
  • Building push-to-talk or hands-free interfaces
The SDK provides energy-based VAD with Silero VAD for accurate speech detection.

​
Automatic VAD with Voice Session

The easiest way to use VAD is through the Voice Agent pipeline: 
final session = await RunAnywhere.startVoiceSession(
  config: VoiceSessionConfig(
    speechThreshold: 0.03,    // Audio level threshold
    silenceDuration: 1.5,     // Seconds of silence before processing
  ),
);

session.events.listen((event) {
  switch (event) {
    case VoiceSessionListening(:final audioLevel):
      // Show audio level meter
      updateAudioMeter(audioLevel);
    case VoiceSessionSpeechStarted():
      print('User started speaking');
    case VoiceSessionProcessing():
      print('Processing speech...');
    default:
      break;
  }
});

​
VoiceSessionConfig for VAD

ParameterTypeDefaultDescription
speechThresholddouble0.03Audio level to trigger speech detection
silenceDurationdouble1.5Seconds of silence before processing
// More sensitive (quieter environments)
final quietConfig = VoiceSessionConfig(
  speechThreshold: 0.01,
  silenceDuration: 1.0,
);

// Less sensitive (noisy environments)
final noisyConfig = VoiceSessionConfig(
  speechThreshold: 0.1,
  silenceDuration: 2.0,
);

​
Building a Voice Level Indicator

class VoiceLevelIndicator extends StatefulWidget {
  final Stream<double> audioLevelStream;

  const VoiceLevelIndicator({required this.audioLevelStream});

  @override
  _VoiceLevelIndicatorState createState() => _VoiceLevelIndicatorState();
}

class _VoiceLevelIndicatorState extends State<VoiceLevelIndicator> {
  double _level = 0.0;
  StreamSubscription? _subscription;

  @override
  void initState() {
    super.initState();
    _subscription = widget.audioLevelStream.listen((level) {
      setState(() => _level = level);
    });
  }

  @override
  Widget build(BuildContext context) {
    return Container(
      height: 10,
      child: LinearProgressIndicator(
        value: _level.clamp(0.0, 1.0),
        backgroundColor: Colors.grey[300],
        valueColor: AlwaysStoppedAnimation(
          _level > 0.03 ? Colors.green : Colors.grey,
        ),
      ),
    );
  }

  @override
  void dispose() {
    _subscription?.cancel();
    super.dispose();
  }
}

​
Voice Session with VAD Events

class VoiceAssistant extends StatefulWidget {
  @override
  _VoiceAssistantState createState() => _VoiceAssistantState();
}

class _VoiceAssistantState extends State<VoiceAssistant> {
  VoiceSessionHandle? _session;
  double _audioLevel = 0.0;
  bool _isSpeaking = false;
  String _status = 'Ready';

  Future<void> _startSession() async {
    _session = await RunAnywhere.startVoiceSession(
      config: VoiceSessionConfig(
        speechThreshold: 0.03,
        silenceDuration: 1.5,
        autoPlayTTS: true,
        continuousMode: true,
      ),
    );

    _session!.events.listen((event) {
      setState(() {
        switch (event) {
          case VoiceSessionListening(:final audioLevel):
            _audioLevel = audioLevel;
            _status = 'Listening...';
          case VoiceSessionSpeechStarted():
            _isSpeaking = true;
            _status = 'Speaking detected';
          case VoiceSessionProcessing():
            _isSpeaking = false;
            _status = 'Processing...';
          case VoiceSessionTranscribed(:final text):
            _status = 'You: $text';
          case VoiceSessionResponded(:final text):
            _status = 'AI: $text';
          case VoiceSessionSpeaking():
            _status = 'AI speaking...';
          case VoiceSessionTurnCompleted():
            _status = 'Ready';
          case VoiceSessionStopped():
            _status = 'Stopped';
          default:
            break;
        }
      });
    });
  }

  void _stopSession() {
    _session?.stop();
  }

  @override
  Widget build(BuildContext context) {
    return Column(
      mainAxisAlignment: MainAxisAlignment.center,
      children: [
        // Audio level indicator
        Container(
          width: 100,
          height: 100,
          decoration: BoxDecoration(
            shape: BoxShape.circle,
            color: _isSpeaking ? Colors.green : Colors.blue,
            boxShadow: [
              BoxShadow(
                color: (_isSpeaking ? Colors.green : Colors.blue)
                    .withValues(alpha: 0.5),
                blurRadius: _audioLevel * 50,
                spreadRadius: _audioLevel * 20,
              ),
            ],
          ),
          child: Icon(
            _isSpeaking ? Icons.mic : Icons.mic_none,
            color: Colors.white,
            size: 48,
          ),
        ),

        SizedBox(height: 24),
        Text(_status, textAlign: TextAlign.center),

        SizedBox(height: 24),
        ElevatedButton(
          onPressed: _session == null ? _startSession : _stopSession,
          child: Text(_session == null ? 'Start' : 'Stop'),
        ),
      ],
    );
  }
}

​
VAD Tuning Tips

Lower thresholds (0.01) detect quieter speech but may trigger on background noise. Higher thresholds (0.1) require louder speech but are more robust to noise.
Shorter duration (0.5s) feels more responsive but may cut off pauses mid-sentence. Longer duration (2.0s) allows natural pauses but feels slower.
Consider measuring ambient noise level and adjusting threshold dynamically.

​
See Also

Voice Agent

Complete voice pipeline

Speech-to-Text

Audio transcription