SenseVoice Transcribe

Transcribe audio to timestamped text using FunASR's iic/SenseVoiceSmall model with fsmn-vad for timestamp anchoring.

Pipeline

1. FSMN-VAD segments audio into speech regions (~258 segments for 30min file)
2. SenseVoice-Small transcribes full audio with merge_vad=True
3. Raw text is split by <|zh|> tags → cleaned via rich_transcription_postprocess()
4. Text segments are proportionally mapped to VAD timestamps
5. Output: [HH:MM:SS] text per line, ~15s granularity

Environment