Voice Activity Detection

The flat VAD surface is three functions on RunAnywhere.

Function	Signature
`detectVoiceActivity`	`(audio: Float32Array, options?: DetectVoiceOptions) => Promise<VADResult>`
`streamVAD`	`(audio: AsyncIterable<Float32Array>, options?: DetectVoiceOptions) => AsyncIterable<VADResult>`
`resetVAD`	`() => boolean`

Audio is Float32Array samples, not PCM16 bytes. Use RunAnywhere.pcm16ToFloat32Samples(arrayBuffer) to convert; it takes an ArrayBuffer of little-endian int16. VAD requires the ONNX backend:

import { ONNX } from '@runanywhere/web-onnx'
await ONNX.register()

Without it, VAD calls fail with ERROR_CODE_FEATURE_NOT_AVAILABLE.

Load a VAD model

const ready = await RunAnywhere.ensureDefaultVAD()

RunAnywhere.defaultVADModelID is 'silero-vad'. ensureDefaultVAD() returns true immediately when a VAD model is already loaded, otherwise it loads the catalogued default. It logs rather than throws on failure, so check the return value. Both flat verbs require a loaded VAD model. Without one they throw SDKException.notInitialized. The energy-threshold detector is a C++ voice-agent fallback; it is not reachable through detectVoiceActivity or streamVAD.

Detect on a buffer

const result = await RunAnywhere.detectVoiceActivity(samples)
if (result.isSpeech) {
  console.log('speech, confidence', result.confidence)
}

VADResult

Field	Type	Meaning
`isSpeech`	`boolean`	Whether the chunk contains speech.
`confidence`	`number`	Model confidence, 0…1.
`energy`	`number`	RMS energy of the chunk.
`durationMs`	`number`	Length of the analysed chunk.
`timestampMs`	`number`	Wall-clock stamp.
`startTimeMs` / `endTimeMs`	`number`	Segment bounds when reported.
`statistics`	`VADStatistics \| undefined`	`currentEnergy`, `currentThreshold`, `ambientLevel`, `recentAvg`.
`errorMessage` / `errorCode`	`string` / `number`	Set on failures.

Options

DetectVoiceOptions extends Partial<VADOptions> and adds two Web-specific fields.

const result = await RunAnywhere.detectVoiceActivity(samples, {
  threshold: 0.5,
  minSpeechDurationMs: 100,
  minSilenceDurationMs: 300,
  maxSpeechDurationMs: 15000,
  includeStatistics: true,
  modelId: 'silero-vad',
  config: { sampleRate: 16000, frameLengthMs: 100 },
})

config is a Partial<VADConfiguration>. Its defaults come from idl/vad_options.proto:

Field	Default
`sampleRate`	`16000`
`frameLengthMs`	`100`
`threshold`	`0.015`
`calibrationMultiplier`	`2.0`
`enableAutoCalibration`	`false`
`windowSizeSamples`	`0` (backend default)
`maxSpeechDurationMs`	`0` (no cap)

VADOptions defaults are threshold: 0 (keep the configured threshold), minSpeechDurationMs: 100, minSilenceDurationMs: 300, maxSpeechDurationMs: 0, and includeStatistics: false.

Streaming VAD only accepts 16 kHz. Passing any other config.sampleRate to streamVAD throws SDKException.invalidConfiguration.

Streaming

async function* micSamples(): AsyncIterable<Float32Array> {
  // yield 16 kHz Float32Array frames from your capture source
}

for await (const result of RunAnywhere.streamVAD(micSamples())) {
  if (result.errorCode !== 0) {
    console.warn('VAD:', result.errorMessage)
    break
  }
  setSpeaking(result.isSpeech)
}

Breaking out of the loop ends the stream.

Capturing audio

AudioCapture from @runanywhere/web/browser wraps getUserMedia and emits frames at the rate the SDK expects.

import { AudioCapture } from '@runanywhere/web/browser'

Handle API

For long-lived sessions where you want to own the component lifecycle, RunAnywhere.vad exposes a handle-based API: create(), configure(handle, config), initialize(handle), loadModel(handle, path, id?, name?), process(handle, samples, options?), statistics(handle), setActivityHandler(handle, handler), start(handle), stop(handle), reset(handle), destroy(handle). setActivityHandler is the one push-style entry point, and it requires a handle:

const handle = RunAnywhere.vad.create()
RunAnywhere.vad.configure(handle, { sampleRate: 16000, frameLengthMs: 100 })
RunAnywhere.vad.initialize(handle)
RunAnywhere.vad.setActivityHandler(handle, (event) => {
  console.log(event.kind)
})
RunAnywhere.vad.start(handle)
// ...
RunAnywhere.vad.stop(handle)
RunAnywhere.vad.destroy(handle)

The flat resetVAD() resets the SDK-owned instance; RunAnywhere.vad.reset(handle) resets a handle you created.

Endpointing an utterance

VAD reports per-chunk state. Turning that into “the user finished speaking” is your job: accumulate frames while isSpeech is true, close the utterance after enough consecutive silent frames, then pass the concatenated samples to RunAnywhere.transcribe(...).

Voice agent

For the full VAD → STT → LLM → TTS turn including microphone ownership, use VoiceAgentMicDriver from @runanywhere/web/browser. See Voice Agent.

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Vibe Coding

Load a VAD model

Detect on a buffer

VADResult

Options

Streaming

Capturing audio

Handle API

Endpointing an utterance

Voice agent

​Load a VAD model

​Detect on a buffer

​VADResult

​Options

​Streaming

​Capturing audio

​Handle API

​Endpointing an utterance

​Voice agent

Load a VAD model

Detect on a buffer

VADResult

Options

Streaming

Capturing audio

Handle API

Endpointing an utterance

Voice agent