Voice activity detection

from wfloat import load_voice_activity_detection

with load_voice_activity_detection("snakers4/silero-vad") as model:
    result = model.detect("recording.wav", return_audio=True)
    for segment in result.segments:
        print(segment.start_ms, segment.end_ms)
        segment.audio.save(f"speech-{segment.id}.wav")

`detect(audio, **options)`

Accepts the same audio inputs as speech to text; arrays need sample_rate. Returns a DetectionResult with segments and stop_reason ("complete" or "cancelled"). Each segment has id, start_ms, and end_ms; return_audio=True adds an Audio clip.

Option	Meaning
`speech_threshold`, `silence_threshold`	Speech entry/exit thresholds.
`min_speech_duration_ms`	Minimum accepted speech duration.
`min_silence_duration_ms`	Silence needed to end a speech segment.
`speech_padding_ms`	Extra audio around detected speech.
`return_audio`	Include clips; default false.
`on_probability`	Receive `probability`, `start_ms`, and `end_ms`.
`cancel_event`	Cooperative cancellation using a `threading.Event`.

Current defaults: speech threshold 0.5, silence threshold 0.15 below it, minimum speech 250 ms, minimum silence 500 ms, padding 30 ms. Thresholds/scores are not interchangeable across models. Padding is clamped to available audio and previous segment boundaries to avoid overlapping clips. Times are relative to supplied audio, not wall-clock time.

`create_session(**options)`

import numpy as np

# With the model still loaded:
session = model.create_session(
    return_audio=True,
    on_speech_end=lambda segment: print(segment.audio),
)
try:
    session.push(np.zeros(16000, dtype=np.float32), sample_rate=16000)
    summary = session.finish()
finally:
    session.cancel()

Sessions process application-supplied audio through synchronous push() and support on_speech_start, on_speech_end, and on_error. One session owns the model. Your application manages capture and buffering.

Live clips are delivered through on_speech_end; the final result contains timing ranges, not every clip. Retain callback audio yourself if needed. finish() flushes the last segment, cancel() preserves completed ranges, and result() retrieves the terminal result. Silence produces no speech segments. Failures raise VadError with partial_result and notify session on_error.

Voice activity detection

import { loadVoiceActivityDetection } from '@wfloat/react-native-wfloat';

const model = await loadVoiceActivityDetection('snakers4/silero-vad');
const audio = { samples: new Float32Array(16000), sampleRate: 16000 }; // One second of silence.
const detection = model.detect(audio, { returnAudio: true });
const result = await detection.result();
console.log(result.segments);
await model.unload();

`detect(audio, options?)`

Accepts the same audio inputs as speech to text. Returns a handle with result() and cancel(). Results contain segments and stopReason: "complete" | "cancelled". A segment has id, startMs, and endMs; returnAudio: true adds its audio.

Option	Meaning
`speechThreshold`, `silenceThreshold`	Speech entry/exit thresholds.
`minSpeechDurationMs`	Minimum accepted speech duration.
`minSilenceDurationMs`	Silence needed to end a speech segment.
`speechPaddingMs`	Extra audio around detected speech.
`returnAudio`	Include completed speech clips; default false.
`onProbability`	Receive `{ probability, startMs, endMs }` model scores.

Current defaults: speech threshold 0.5, silence threshold 0.15 below it, minimum speech 250 ms, minimum silence 500 ms, padding 30 ms. Scores and threshold effectiveness are not interchangeable across models. Padding is clamped to available audio and previous segment boundaries to avoid overlapping clips.

`createSession(options?)`

// With the model still loaded:
const session = await model.createSession({
  returnAudio: true,
  onSpeechStart: event => console.log('Speech started', event.startMs),
  onSpeechEnd: event => console.log(event.audio),
  onError: error => console.error(error),
});
await session.startMicrophone(); // From a user action.
// Later:
const summary = await session.finish();

Sessions also support push({ samples, sampleRate }), attachMicrophone(source), result(), and cancel(). finish() ends input and flushes the last segment; result() only waits. One session owns the model at a time.

Live onSpeechEnd delivers clips when returnAudio is enabled. The final session result retains timing ranges, not all audio clips; keep callback clips yourself if needed. Times are relative to the supplied audio/session. Silence produces no speech segments. Failures use VadError.partialResult and session onError.

See speech to text for shared microphone setup and platform-specific capture behavior.

Voice activity detection

import { loadVoiceActivityDetection } from '@wfloat/wfloat-web';

const model = await loadVoiceActivityDetection('snakers4/silero-vad');
const audio = { samples: new Float32Array(16000), sampleRate: 16000 }; // One second of silence.
const detection = model.detect(audio, { returnAudio: true });
const result = await detection.result();
console.log(result.segments);
await model.unload();

`detect(audio, options?)`

Accepts the same audio inputs as speech to text. Returns a handle with result() and cancel(). Results contain segments and stopReason: "complete" | "cancelled". A segment has id, startMs, and endMs; returnAudio: true adds its audio.

Option	Meaning
`speechThreshold`, `silenceThreshold`	Speech entry/exit thresholds.
`minSpeechDurationMs`	Minimum accepted speech duration.
`minSilenceDurationMs`	Silence needed to end a speech segment.
`speechPaddingMs`	Extra audio around detected speech.
`returnAudio`	Include completed speech clips; default false.
`onProbability`	Receive `{ probability, startMs, endMs }` model scores.

Current defaults: speech threshold 0.5, silence threshold 0.15 below it, minimum speech 250 ms, minimum silence 500 ms, padding 30 ms. Scores and threshold effectiveness are not interchangeable across models. Padding is clamped to available audio and previous segment boundaries to avoid overlapping clips.

`createSession(options?)`

// With the model still loaded:
const session = await model.createSession({
  returnAudio: true,
  onSpeechStart: event => console.log('Speech started', event.startMs),
  onSpeechEnd: event => console.log(event.audio),
  onError: error => console.error(error),
});
await session.startMicrophone(); // From a user action.
// Later:
const summary = await session.finish();

Sessions also support push({ samples, sampleRate }), attachMicrophone(source), result(), and cancel(). finish() ends input and flushes the last segment; result() only waits. One session owns the model at a time.

Live onSpeechEnd delivers clips when returnAudio is enabled. The final session result retains timing ranges, not all audio clips; keep callback clips yourself if needed. Times are relative to the supplied audio/session. Silence produces no speech segments. Failures use VadError.partialResult and session onError.

See speech to text for shared microphone setup and platform-specific capture behavior.