Detect voice activity using Silero VAD

Use This When

• Building voice assistants that should only process speech segments to save compute
• Implementing push-to-talk alternatives with automatic speech detection
• Reducing ASR and LLM costs by filtering out silence and non-speech audio
• Creating audio recording systems that skip silent intervals

What It Does

• Detects presence of human speech in audio frames using lightweight Silero VAD model
• Resamples audio to 16kHz and returns boolean indicating speech presence
• Generates timestamp-based speech segments for temporal speech localization
• Runs efficient neural model suitable for real-time streaming applications

Works Best With

• Audio inputs → this component → transcribe-audio to gate ASR on speech segments only
• Integration with denoise-audio → this component → ASR for clean speech detection
• Voice assistant pipelines where VAD triggers wake word detection or command processing
• Recording systems that need automatic silence removal or speech-only archival

Caveats

• Very low SNR or loud music can cause false positive speech detections
• Model optimized for 16kHz; quality degrades if original recording is lower sample rate
• Frame-level decisions may miss very short utterances shorter than analysis window
• Background babble or TV dialogue may trigger false positives in multi-speaker environments