Transcribe

/transcribe — Local audio transcription with mlx-whisper

Transcribe audio files (e.g. iPhone voice memos sent via AirDrop) with high accuracy, entirely on-device. Audio never leaves the Mac — safe for confidential recordings.

Requirements: Apple Silicon Mac, mlx-whisper (pip install mlx-whisper), ffmpeg (brew install ffmpeg).

Steps

1. Find the audio file

If the user gave a file path, use it. Otherwise look for recent audio in Downloads:

find ~/Downloads -maxdepth 1 \( -iname "*.m4a" -o -iname "*.mp3" -o -iname "*.wav" -o -iname "*.aac" -o -iname "*.aiff" -o -iname "*.caf" \) -mtime -2 -exec ls -lt {} +

If there are multiple candidates, pick the newest and confirm the file name and size with the user.

2. Check the duration

ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "<file>"

Rough guide: on an M-series Mac, processing takes about 1/20 of the recording length (80 min → ~4 min).

3. Confirm the language (important)

Never hard-code --language. Forcing the wrong language (e.g. ja on an English meeting) can send Whisper into a full-file hallucination loop.

• If the user stated the language, pass it (--language en / --language ja)
• If unknown, omit the flag and let Whisper auto-detect

4. Run the transcription

Locate the CLI first — it is often installed outside PATH:

command -v mlx_whisper || ls ~/Library/Python/*/bin/mlx_whisper ~/.local/bin/mlx_whisper 2>/dev/null

For recordings longer than ~10 minutes, run in the background:

mlx_whisper "<file>" \
  --model mlx-community/whisper-large-v3-turbo \
  --condition-on-previous-text False \
  --output-format all --output-name "<basename>" \
  --output-dir <temp-dir> --verbose False

• Always pass --condition-on-previous-text False — it prevents repetition loops
• First run downloads the model (~1.6 GB) to the Hugging Face cache

5. Quality check (mandatory)

sort "<output>.txt" | uniq -c | sort -rn | head -5

• Repeated short backchannels ("Okay.", "Yeah.") are normal
• The same long sentence repeated dozens of times = hallucination. Suspect a wrong language flag; cut a 60-second sample and retry with auto-detect:

  ffmpeg -y -v error -ss 600 -t 60 -i "<file>" -ac 1 -ar 16000 sample.wav
  

6. Deliver the output

Copy the results somewhere convenient for the user (e.g. Downloads) and reveal them in Finder:

cp <output>.txt ~/Downloads/"<original-name>_transcript.txt"
cp <output>.srt ~/Downloads/"<original-name>_transcript.srt"
open -R ~/Downloads/"<original-name>_transcript.txt"

To send back to an iPhone: right-click → Share → AirDrop (the final send click cannot be automated — macOS limitation). Avoid ad-hoc HTTP servers or unapproved cloud uploads for confidential recordings.

7. Summarize the content

Read the full transcript and report back:

• Type of meeting/recording and main topics
• Decisions, open questions, next actions
• Notable statements (numbers, risks, follow-ups)

Offer to reformat as meeting minutes or export to PDF if useful.

Notes

• Model: whisper-large-v3-turbo — noticeably more accurate than the built-in transcription on Teams or stock Android/iOS voice apps
• mlx-whisper uses Apple's MLX framework, so it only runs on Apple Silicon (M1 or later)