WebRTC Voice Activity Detector

Warnings

• gotcha The WebRTC VAD has strict audio input requirements: 16-bit, mono PCM audio, sampled at 8000, 16000, 32000, or 48000 Hz. Frames must be exactly 10, 20, or 30 ms in duration.
  Fix: Ensure your audio input (e.g., from a microphone or file) is pre-processed to match these specifications before passing it to `vad.is_speech()`.
• gotcha The `webrtcvad` package (wiseman/py-webrtcvad) can be difficult to install on some platforms due to its C/C++ dependencies and lack of pre-built wheels for all Python versions/OS combinations.
  Fix: Consider using `pip install webrtcvad-wheels` instead, which is a fork specifically designed to provide pre-compiled binary wheels for easier installation across various platforms. The API remains the same. If sticking with `webrtcvad`, ensure you have a C/C++ compiler installed and up-to-date Python development headers.
• gotcha Version 2.0.10 fixed a memory leak in the `is_speech()` method. While 2.0.10 should contain this fix, later versions of the `webrtcvad-wheels` fork (e.g., 2.0.13) address further memory leak issues.
  Fix: Ensure you are using at least `webrtcvad` 2.0.10. For the most robust memory handling, consider migrating to `webrtcvad-wheels` and using its latest version.
• gotcha The WebRTC VAD is a simple, real-time oriented model and may produce false positives for non-speech sounds (e.g., music, birdsong) or false negatives in very noisy environments, even at high aggressiveness settings.
  Fix: For applications requiring higher accuracy in challenging audio environments or more nuanced classification beyond simple speech/no-speech, consider integrating more advanced VAD algorithms or machine learning models.

Install

• pip install webrtcvad Install stable version

Imports

• Vad

  import webrtcvad
  vad = webrtcvad.Vad()

Quickstart

Initializes the VAD, sets its aggressiveness mode, and demonstrates classifying silence and a mock speech segment. It highlights the strict audio format requirements: 16-bit mono PCM at 8000, 16000, 32000, or 48000 Hz, with frame durations of 10, 20, or 30 ms.

import webrtcvad
import struct

# WebRTC VAD requires 16-bit mono PCM audio at specific sample rates
# and frame durations (10, 20, or 30 ms).
sample_rate = 16000 # Hz
frame_duration_ms = 30 # ms
bytes_per_sample = 2 # 16-bit audio

# Calculate frame size in bytes
frame_size_bytes = int(sample_rate * (frame_duration_ms / 1000.0) * bytes_per_sample)

# Create a VAD instance with an aggressiveness mode (0-3)
# 0: least aggressive, 3: most aggressive
vad = webrtcvad.Vad(3)

# Create a silent audio frame (16-bit mono PCM)
silence_frame = b'\x00\x00' * int(frame_size_bytes / bytes_per_sample)

# Create a mock speech-like frame (simple sine wave for demonstration)
# In a real application, this would come from an audio input.
speech_frame = b''
for i in range(int(frame_size_bytes / bytes_per_sample)):
    # Simple sine wave approximation for a speech-like signal
    amplitude = 10000 # Max 32767 for 16-bit
    value = int(amplitude * (i % 30 < 15) - amplitude * (i % 30 >= 15)) # Square wave approximation
    speech_frame += struct.pack('<h', value)


print(f"Processing frame of {frame_duration_ms} ms at {sample_rate} Hz")

# Test with silence
is_speech_silence = vad.is_speech(silence_frame, sample_rate)
print(f"Silence frame contains speech: {is_speech_silence}")

# Test with speech-like audio
is_speech_mock = vad.is_speech(speech_frame, sample_rate)
print(f"Mock speech frame contains speech: {is_speech_mock}")

# You can also set the mode after initialization
vad.set_mode(1)
print(f"VAD aggressiveness set to 1.")