Voice Activity Detection

the system continuously listens to the mic, but only starts ‘recording’ when it detects speech, and stops when there’s silence for a certain duration.

Without VAD:

• Wasted CPU/GPU transcribing long silences.
• Context window gets filled with “um… yeah… (5 seconds of silence)” instead of meaningful content.
• Risk of producing huge files and high latency in processing.

With VAD:

• The bot reacts only to speech.
• Silences become natural ‘breaks’ for sending content to the LLM.
• You can better preserve meeting flow without constant interruptions.

can use a VAD library