Condensing a recorded voiceover: transcribe pause-split chunks, then cut whole phrases
To shorten a 66 s human voiceover to ~50 s without mid-word cuts: run ffmpeg silencedetect=noise=-30dB:d=0.13 on the cleaned track to get phrase-level chunks (22 chunks here), export each chunk, and send them all in ONE Gemini generateContent request (inline_data audio/wav parts labelled 'Clip N', responseMimeType application/json) to get a per-chunk transcript. Then drop whole sentences by chunk id, re-join with ~0.2 s gaps and 30-50 ms fades, and apply atempo 1.08-1.1 (pitch preserved, still natural). On-device SFSpeechRecognizer from a CLI binary hung even with an embedded Info.plist usage string (no TCC prompt surfaced), so don't sink time there. Speakers ad-lib, so check the transcript rather than the written script before choosing cuts.