Transcribe

6. Extract audio with ffmpeg:

   ffmpeg -y -i /tmp/transcribe-work/video.mp4 -vn -acodec pcm_s16le -ar 16000 -ac 1 /tmp/transcribe-work/audio.wav
   

7. Delete the video file to save disk space: rm /tmp/transcribe-work/video.mp4.

Transcribe

8. Check if mlx_whisper is importable in Python 3. If not, install it: pip3 install mlx-whisper.
9. Run transcription with the following Python script:

   import mlx_whisper
   result = mlx_whisper.transcribe(
       '/tmp/transcribe-work/audio.wav',
       path_or_hf_repo='mlx-community/whisper-small-mlx',
       language='en'
   )
   with open('/tmp/transcribe-work/transcript.txt', 'w') as f:
       for seg in result['segments']:
           start = int(seg['start'])
           m, s = divmod(start, 60)
           f.write(f'[{m:02d}:{s:02d}] {seg["text"].strip()}\n')
   

   • If disk space is tight (the large model fails), fall back to mlx-community/whisper-small-mlx.
   • For non-English content, omit the language parameter or set it appropriately.

Summarize

10. Read the full transcript from /tmp/transcribe-work/transcript.txt.
11. Produce a structured summary with:
    • Title and metadata (speakers, host, source)
    • Key takeaways — the 4-6 most important points, each with a bold heading and 2-3 sentence explanation
    • Notable quotes or claims if any stand out
    • Keep the summary concise but substantive. Match the depth to the content length (short video = brief summary, long podcast = detailed breakdown).
12. Present the summary to the user. Mention that the full timestamped transcript is at /tmp/transcribe-work/transcript.txt.

Cleanup

13. Delete /tmp/transcribe-work/audio.wav after transcription to free space. Keep transcript.txt.