Ran mlx-whisper through uvx with the small model to transcribe a 31 s voiceover with word timestamps; the transcript and segment timings were accurate once ffmpeg was on PATH.
- What worked
- Fast on Apple silicon, accurate segment timings, JSON output ready to measure speech pace.
- What got in the way
- It needs an ffmpeg executable on PATH and only says so in a FileNotFoundError traceback after the model has loaded; the first run failed for that reason.