RealtimeTenVad (TEN VAD – Realtime, Node.js)

Tiny realtime VAD wrapper around a WASM backend. Feed it 16 kHz mono float audio (Float32Array in [-1,1]), and get start/end callbacks for detected speech.

Adapted from TEN-VAD.

Installation and Usage

npm install realtime-ten-vad

Require and use as follows:

// Use main export (index.js):
const { RealtimeTenVad } = require('realtime-ten-vad');

const vad = await RealtimeTenVad.new({
  // Callbacks
  onSpeechStart() { /* speech just began */ },
  onSpeechEnd(seg) { /* seg: Float32Array @ 16 kHz with pre/post padding */ },
  onVADMisfire() { /* segment shorter than minSpeechMs */ },

  // Tunables (all optional; shown with defaults)
  positiveSpeechThreshold: 0.6,
  negativeSpeechThreshold: 0.4,
  minSpeechFrames: 3,
  minSilenceFrames: 8,
  probSmoothing: 0.2,
  energyGateDb: null,
  preEmphasis: 0.0,
  preSpeechPadMs: 80,
  postSpeechPadMs: 160,
  minSpeechMs: 150,
});

// stream audio
vad.processAudio(float32Chunk);
// when you're done:
vad.flush();
vad.destroy();

Options (what they do)

Callbacks

• onSpeechStart(): void Fired once when the VAD transitions into speech (after the minimum entry condition is met). No arguments.
• onSpeechEnd(seg: Float32Array): void Fired when a speech segment is finalized (after trailing silence + post-pad). seg is 16 kHz mono, includes configured pre/post padding.
• onVADMisfire(): void Fired instead of onSpeechEnd if the detected segment’s duration is < minSpeechMs.

Tunables / knobs

| Option | Type (range) | Default | Meaning & tips | |-------------------------- | -------------------:| -------:| -------------------------------------------------------------------------------------------------------------------- | | positiveSpeechThreshold | number (0..1) | 0.6 | Probability threshold to enter speech when in silence. Higher = stricter start | | negativeSpeechThreshold | number (0..1) | 0.4 | Probability threshold in speech, keeps speech “latched” until confidence < this value. Lower = more tolerant. | | (hysteresis) | — | — | If negative > positive, values are automatically swapped to ensure correct behavior | | minSpeechFrames | number (frames) | 3 | Minimum consecutive voice frames (each 256 samples ≈ 16 ms) needed for speech start. 3 ≈ 48 ms. | | minSilenceFrames | number (frames) | 8 | Minimum consecutive non-voice frames to consider speech ended | | probSmoothing | number (0..1) | 0.2 | EMA smoothing of VAD probability. 0 = reacts instantly; →1 = slow. | | energyGateDb | number \| null | null | RMS energy gate in dBFS. Frames below this are treated as silence; null disables. | | preEmphasis | number (~0..0.97) | 0.0 | Pre-emphasis for VAD only. Does not affect callback audio. | | preSpeechPadMs | number (ms) | 80 | Leading context included at the start of each segment | | postSpeechPadMs | number (ms) | 160 | Trailing context appended after speech ends (collected during silence) prior to segment finalize | | minSpeechMs | number (ms) | 150 | Segments shorter than this are considered misfires → onVADMisfire |

Notes & expectations

• Input: processAudio(Float32Array) at 16 kHz mono, range [-1, 1]. Anything else throws.
• Framing: Fixed hop of 256 samples (≈16 ms) internally.
• System deps: For microphone recording, you may require platform tools (Linux: arecord/sox, macOS: sox).
• Lifecycle: Call flush() to close in-progress segments (adds post-pad) before destroy().
• CLI demo/test: See test.js and record.js. System microphone permissions/tools required.