Voice Activity Detection

1. pip install silero-vad (Silero VAD — the gold-standard open-weight voice activity detection, ~2 MB).
2. Python script for real-time VAD:

import torch, pyaudio
model, utils = torch.hub.load("snakers4/silero-vad", "silero_vad", trust_repo=True)
(get_speech_timestamps, _, _, _, _) = utils

CHUNK = 512  # 32ms at 16kHz
audio = pyaudio.PyAudio()
stream = audio.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True, frames_per_buffer=CHUNK)
while True:
    chunk = stream.read(CHUNK)
    tensor = torch.frombuffer(chunk, dtype=torch.int16).float() / 32768.0
    speech_prob = model(tensor, 16000).item()
    if speech_prob > 0.5:
        print(f"Speech detected: {speech_prob:.2f}")

1. First detection in <10 ms — Silero VAD is real-time on CPU, 50× faster than real-time in batch mode.
2. For file-based VAD: pip install silero-vad + from silero_vad import read_audio, get_speech_timestamps → returns start/end times of all speech segments in an audio file.
3. Use cases: audio pre-processing before STT (skip silence), real-time voice activity for chatbots, meeting recording segmentation, audio stream monitoring.

VAD is the lightest AI task on this list. Silero VAD (~2 MB) runs at 500-1000× real-time on any laptop CPU. A 1-hour audio file VAD-processes in ~3-6 seconds. No GPU needed. Even a Raspberry Pi 4 ($35) runs Silero VAD in real-time. Any $200 Chromebook handles production VAD. VAD is so computationally trivial that hardware is never the bottleneck — code integration is the only challenge. If you can run Python, you can run VAD. Total cost to add VAD to any project: $0 hardware, 5 lines of Python.

VAD doesn't need "serious" hardware. The model is 2 MB and runs on CPU at 500-1000× real-time. For production audio pipelines processing 1000+ concurrent streams (call center monitoring, broadcast VAD), any modern server CPU (AMD EPYC, Intel Xeon) handles it trivially. The scaling challenge is I/O and audio streaming infrastructure, not compute. Budget for audio routing (Kafka streams, WebRTC) and storage, not GPU. A $500 refurbished server with 32 cores handles 10,000+ simultaneous VAD streams. VAD is the poster child for "you don't need a GPU for AI."

The mistake: Using VAD with default threshold (0.5) and wondering why half the speech is cut off or why background noise triggers false positives. Why it fails: VAD thresholds are environment-dependent. In a quiet studio, speech_prob=0.1 reliably separates speech from silence. In a noisy café, speech_prob=0.7 might be the minimum to avoid false triggers on background chatter. The default threshold is tuned for clean speech in quiet environments. The fix: Tune the threshold for your environment. Record 1 minute of ambient noise + 1 minute of target speech. Plot speech_prob over time. Set the threshold above the highest noise probability and below the lowest speech probability. For dynamic environments (varying noise levels), use Silero VAD's built-in adaptive thresholding or pair with a noise estimator. A one-size threshold fits no environment.