Silero VAD

Decibri ships Silero VAD v5 bundled. Set the VAD mode to 'silero' to enable neural voice activity detection with no extra dependencies, no separate model download, and no sherpa-onnx required. This page covers both Python and Node.js. Tabs above each code block switch between languages.

What this does

Silero v5 is a neural voice activity detection model. It runs locally in Rust via ONNX Runtime. No cloud API, no server, no fine-tuning required. The ~2.3 MB ONNX model ships inside the decibri package on both PyPI and npm.

Decibri offers two VAD modes out of the box: an RMS energy detector ('energy') that uses simple amplitude thresholding, and the Silero v5 neural model ('silero') that handles noisy environments, background music, and multi-speaker scenarios with higher accuracy. The model set is exactly these two.

Prerequisites

Install Decibri

The Silero v5 ONNX model is bundled in both the PyPI and npm packages. No separate download, no sherpa-onnx needed. The install command above switches with the language tabs in the code blocks below.

Code walkthrough

1. Create a Microphone with Silero VAD

Set the VAD mode to 'silero' to enable voice activity detection with the neural model instead of the RMS energy detector. The bare string shorthand uses the mode defaults: a 0.5 probability threshold for Silero and a 300 ms holdoff.

import decibri

mic = decibri.Microphone(sample_rate=16000, channels=1, vad="silero")

const { Microphone } = require('decibri');

const mic = new Microphone({ sampleRate: 16000, channels: 1, vad: 'silero' });

To tune the detection threshold or the holdoff, pass a Vad config object instead of the bare string. In Python that is decibri.Vad(model=, threshold=, holdoff_ms=); in Node.js it is a { model, threshold, holdoffMs } object. The 'energy' mode is configured the same way.

import decibri

# Override the threshold and holdoff with a Vad config object.
mic = decibri.Microphone(
    sample_rate=16000,
    channels=1,
    vad=decibri.Vad(model="silero", threshold=0.6, holdoff_ms=200),
)

const { Microphone } = require('decibri');

// Override the threshold and holdoff with a Vad config object.
const mic = new Microphone({
  sampleRate: 16000,
  channels: 1,
  vad: { model: 'silero', threshold: 0.6, holdoffMs: 200 },
});

2. React to speech and silence

The two bindings expose voice activity differently. In Node.js, decibri emits a 'speech' event when voice activity starts and a 'silence' event when audio stays below the threshold for the holdoff window (the holdoffMs field, default 300 ms); read the latest score from mic.vadScore. In Python there are no events: read mic.is_speaking and mic.vad_score as properties in the chunk loop, and detect the transitions yourself. Both patterns ride alongside the regular audio stream.

prev_speaking = False

with mic:
    for chunk in mic:
        if mic.is_speaking and not prev_speaking:
            print("[speech started]")
        elif not mic.is_speaking and prev_speaking:
            print("[speech ended]")
        prev_speaking = mic.is_speaking
        # chunk is int16 PCM bytes; mic.vad_score is the latest Silero probability

mic.on('speech', () => {
  console.log('[speech started]');
});

mic.on('silence', () => {
  console.log('[speech ended]');
});

mic.on('data', (chunk) => {
  // Audio arrives continuously. mic.vadScore is the latest Silero probability.
});

3. Clean shutdown

Stop the microphone when the user presses Ctrl+C. In Python the with block stops capture on exit, so catching KeyboardInterrupt is enough. In Node.js, call mic.stop() from a SIGINT handler.

try:
    with mic:
        for chunk in mic:
            ...  # detect speech as shown above
except KeyboardInterrupt:
    print("\nStopping...")

process.on('SIGINT', () => {
  console.log('\nStopping...');
  mic.stop();
  process.exit(0);
});

Full example

import decibri

mic = decibri.Microphone(sample_rate=16000, channels=1, vad="silero")

print("Listening for speech... (Ctrl+C to stop)\n")
prev_speaking = False
try:
    with mic:
        for chunk in mic:
            if mic.is_speaking and not prev_speaking:
                print("[speech start]")
            elif not mic.is_speaking and prev_speaking:
                print("[speech end]")
            prev_speaking = mic.is_speaking
            # mic.vad_score is the latest Silero probability (0.0 to 1.0)
except KeyboardInterrupt:
    print("\nStopping...")

'use strict';

const { Microphone } = require('decibri');

const mic = new Microphone({ sampleRate: 16000, channels: 1, vad: 'silero' });

mic.on('speech', () => console.log('[speech start]'));
mic.on('silence', () => console.log('[speech end]'));

mic.on('error', (err) => console.error('Mic error:', err.message));

process.on('SIGINT', () => {
  console.log('\nStopping...');
  mic.stop();
  process.exit(0);
});

console.log('Listening for speech... (Ctrl+C to stop)\n');

Configuration options

Voice activity detection is off unless you enable it. There are two ways to turn it on.

mic = decibri.Microphone(vad="silero")   # or "energy"

mic = decibri.Microphone(
    vad=decibri.Vad(model="silero", threshold=0.6, holdoff_ms=200)
)

const mic = new Microphone({ vad: 'silero' });   // or 'energy'

const mic = new Microphone({
  vad: { model: 'silero', threshold: 0.6, holdoffMs: 200 }
});