Voice Activity Detection¶

Voice Activity Detection (VAD) enables automatic transcription when you stop speaking.

Enable VAD¶

asr2clip --vad

With VAD enabled, transcription triggers when:

1. Speech is detected (audio above threshold)
2. Followed by silence (audio below threshold for the specified duration)

How It Works¶

asr2clip uses multi-feature detection for robust speech recognition:

• RMS energy — measures overall audio volume
• Zero-crossing rate — distinguishes speech from noise (speech has lower ZCR)
• Speech-band frequency ratio — checks energy concentration in 300-3000 Hz (the human speech band)

All three features must agree before audio is classified as speech, reducing false triggers from keyboard noise, fans, or other non-speech sounds.

Adaptive Threshold¶

By default, VAD uses adaptive threshold that adjusts to ambient noise in real-time:

# Adaptive is enabled by default with --vad
asr2clip --vad

# Disable adaptive threshold (use fixed value)
asr2clip --vad --no_adaptive

The adaptive threshold continuously monitors background noise and adjusts sensitivity accordingly, so you don't need to recalibrate when the environment changes.

Calibration¶

Measure your environment's ambient noise to set an appropriate threshold:

asr2clip --calibrate

This records a short sample of ambient noise and suggests a threshold value.

Custom Settings¶

# Custom silence threshold and duration
asr2clip --vad --silence_threshold 0.005 --silence_duration 2.0

VAD Options¶

Behavior Details¶

• Minimum speech duration: 0.5s — short bursts of noise are ignored
• Maximum speech duration: 30s — transcription is forced after 30 seconds of continuous speech
• Speech gap tolerance: 0.3s — brief pauses within speech are not treated as silence
• Auto-calibrate on startup: ambient noise level is measured when recording begins