Voice Activity Detection - RunAnywhere Documentation

RunAnywhere.vad answers “is someone talking right now”. It is what a voice interface uses to decide when a turn has ended.

val result = RunAnywhere.vad.detect(AudioInput.file(path))

Signatures

suspend fun detect(audio: AudioInput, options: VadOptions? = null): VadResult
fun openStream(format: AudioFormatSpec, options: VadOptions? = null): VadStream
fun detectStream(audio: Flow<AudioInput>, options: VadOptions? = null): Flow<VadEvent>

Streaming from a microphone

val stream = RunAnywhere.vad.openStream(
    AudioFormatSpec(AudioEncoding.PcmF32Le, 16000, 1)
)

// per buffer
stream.pushFrame(AudioFrame(bytes, count))

// when done
stream.flush()
stream.finish()
stream.close()

Streaming from a Flow

RunAnywhere.vad.detectStream(audioFlow).collect { event ->
    when (event) {
        is VadEvent.SpeechStart -> isSpeaking = true
        is VadEvent.SpeechEnd -> isSpeaking = false
        else -> {}
    }
}

Options

val options = VadOptions(
    activationThreshold = 0.5f,
    minSpeechMs = 250,
    minSilenceMs = 500,
    prefixPaddingMs = 300
)

Field	Default	Meaning
`activationThreshold`	model default	How confident before it counts as speech
`minSpeechMs`	`DEFAULT_MIN_SPEECH_MS`	Ignore bursts shorter than this
`minSilenceMs`	`DEFAULT_MIN_SILENCE_MS`	Silence before a turn is over
`prefixPaddingMs`	proto default	Audio kept from before speech started

minSilenceMs is the one to tune. Too short and the assistant interrupts someone who paused to think; too long and it feels sluggish. prefixPaddingMs matters because detection lags the first syllable. Without padding, “hello” reaches the transcriber as “ello”.

You may not need this

For a spoken conversation, use the voice session. It runs VAD, transcription, generation, and speech together, with turn-taking and interruption already wired up.