Transcribe

Local speech-to-text with Whisper. On Apple Silicon, mlx-whisper runs on the GPU and is fast (a 15-20 minute file takes a couple of minutes). Elsewhere, openai-whisper does the same job more slowly. Both run through uvx, so nothing is installed permanently.

Steps

1. Locate the input. Use the path the user gave, or the path printed by the youtube-search download step (--print after_move:filepath). Done when ls confirms the file exists.

2. Convert only if needed. Both engines read m4a, mp3, wav, flac, ogg, webm and mp4 directly via ffmpeg. Convert first with the ffmpeg-cli skill (from this marketplace's ffmpeg plugin) when the file is an unusual format (.amr, .wma, .caf), or an .opus voice note, which decodes more reliably as m4a:

   ffmpeg -i "input.opus" -c:a aac -b:a 128k "input.m4a"
   

3. Pick the language. Pass --language <code> (en, es, fr, ...) whenever the user states it or the context makes it obvious - auto-detection costs time and can misidentify short clips. Omit the flag only when the language is genuinely unknown.

4. Pick the engine with uname -sm: Darwin arm64 runs mlx-whisper; anything else runs openai-whisper.

5. Run it (allow up to 600 s; the first run also downloads the model, about 1.5 GB):

   Apple Silicon - mlx-whisper