8.7.13. dasLLAMA-12 — Voice Activity Detection

Voice-activity detection answers one question about audio: where is the speech? dasLLAMA ships Silero-VAD with its weights checked in beside the module, and a four-second public-domain speech fixture next to them — so this is the one dasLLAMA tutorial that runs with no downloads and no arguments:

daslang.exe -jit tutorials/dasLLAMA/12_voice_activity.das

Pass your own 16 kHz mono PCM16 WAV to analyze it instead; pass an ASR model (as in tutorial 07) and the last section transcribes what the gate lets through.

8.7.13.1. A trained gate, not an energy gate

vad_probs runs the model over 512-sample chunks — 32 ms at 16 kHz — and returns one speech probability per chunk:

let probs <- vad_probs(m, speech)

The tutorial feeds it three loud fakes — white noise, a pure tone, and a formant-shaped fake vowel — plus real speech. An energy detector fires on all three fakes. Silero pins noise and tone near zero; the fake vowel, built to mimic speech’s structure, brushes the 0.5 threshold; real speech pins the ceiling:

peak speech probability: fake vowel 0.62, noise 0.01, tone 0.004
peak speech probability: real speech 0.9997

The gate reads structure, not volume. One practical note: the model is stateful — it judges each chunk in the context of what came before, so synthetic tests give every span some leading silence.

8.7.13.2. Segments: where the speech is

find_speech_timestamps turns the probability curve into [start, end) sample regions with silero’s hysteresis: a segment opens at probability ≥ threshold (0.5), closes only after it stays below neg_threshold (0.35) for min_silence (100 ms), segments shorter than min_speech (250 ms) are dropped, and boundaries get speech_pad (30 ms) of margin. default_vad_opts() is that tuning; override fields for a different trade:

var segs <- find_speech_timestamps(m, speech, default_vad_opts())
for (g in segs) {
    print("{float(g.start) / 16000.0} - {float(g.end) / 16000.0} s\n")
}
// -> 0.322 - 2.27 s
//    3.266 - 4 s

8.7.13.3. Streaming: the live-microphone shape

A live capture cannot wait for the whole recording. VadIter is the online segmenter: feed audio in whatever chunk sizes arrive, and it emits a speech_start / speech_end event the moment each boundary is crossed; leftover samples buffer until the next call. Same hysteresis as above:

var it = VadIter()
vad_iter_reset(it, default_vad_opts())
vad_iter_feed(m, it, speech) $(ev) {
    print("{ev.kind} at {float(ev.sample) / 16000.0} s\n")
}

8.7.13.4. The pipeline payoff

VAD in front of a speech-to-text model is the standard rig: long silences never reach the transcriber, and each detected segment becomes its own transcribe call. The tutorial’s last section does exactly that when an ASR model is on the command line, and skips politely when one is not — tutorial 07 covers the ASR surface itself.