Silero VAD Production Deployment

A voice activity detector is the quiet workhorse of any real-time audio pipeline. Silero VAD runs in milliseconds, decides whether each audio chunk contains speech, and is how you avoid transcribing endless silence on dedicated GPU hosting.

Contents

• Why VAD matters
• Deployment
• Streaming pattern
• Threshold tuning

Why VAD

Without VAD, an always-on microphone feeds silence, background noise, and speech alike to Whisper. Whisper hallucinates on silence. VAD gates the transcriber – only speech-containing chunks go through. Benefits:

• No hallucinated transcripts during silence
• Lower compute cost (Whisper only runs on real speech)
• Accurate speech/silence boundaries for downstream tasks

Deployment

import torch

model, utils = torch.hub.load(
    repo_or_dir="snakers4/silero-vad",
    model="silero_vad",
    force_reload=False,
)
(get_speech_timestamps, _, _, _, _) = utils

audio = read_audio("input.wav")
timestamps = get_speech_timestamps(audio, model, threshold=0.5)

Silero VAD is tiny (~2 MB). Runs on CPU or GPU. GPU gives a minor speed-up but is not required.

Streaming

For real-time audio, feed Silero VAD 512-sample chunks (32 ms at 16 kHz) as they arrive. Accumulate chunks flagged as speech, and when you get 2+ seconds of speech-then-silence, hand the buffer to Whisper:

while True:
    chunk = read_from_mic()
    is_speech = vad(chunk) > 0.5
    if is_speech:
        buffer.append(chunk)
    elif buffer and silence_count > 30:  # ~1 second of silence
        transcribe(buffer)
        buffer.clear()

Threshold

Default threshold 0.5 works for clean audio. In noisy environments raise to 0.6-0.7 to reduce false positives. In quiet audio with soft speakers drop to 0.3-0.4 to catch quiet speech.

See Whisper Turbo and Whisper + diarization.