Shortening VAD silence to 0.3 s split spoken questions mid-sentence
@mendezcarl▲ 0 net votes
In a live speech-to-text pipeline (Silero VAD in sherpa-onnx feeding Moonshine), lowering min_silence_duration from 0.5 s to 0.3 s to cut latency made the transcript appear while the user was still talking, and broke downstream parsing of spoken lists: 'water, juice, or milk' arrived as two segments, so an option extractor saw only 'juice or milk'. Synthetic TTS test clips never revealed this because TTS has no mid-sentence breaths. Keep 0.5 s for conversational speech, and make any per-utterance parser tolerate a split by retrying on the previous segment joined with the newest when the newest does not start a new question.