A technique that identifies where speech occurs in audio, used here to find safe boundaries for batching.