Audio Transcription

Use this skill whenever the user asks to transcribe an audio/video file, a Voice Memos export, dictation, lecture, meeting recording, or "bad audio".

Core rules

1. Preserve temporary inputs immediately. Voice Memo share-sheet paths under ~/Library/Containers/com.apple.VoiceMemos/Data/tmp/.com.apple.uikit.itemprovider... can disappear. Before probing or experimenting, copy the file to stable /private/tmp/audio-transcription-inputs/.
2. Use cached local models, not cloud APIs. Prefer MLX Whisper via uvx --from mlx-whisper mlx_whisper; Hugging Face models must be cached in ~/.cache/huggingface/hub/.
3. Force language when known. For Armin's own dictations this is usually English with an Austrian/German accent, even when the filename is German. Do not infer language from filename alone.
4. For bad audio, run a hallucination-resistant pass. Use --condition-on-previous-text False, --word-timestamps True, and --hallucination-silence-threshold 2.
5. Deliver a cleaned best-effort transcript. Compare model output with timestamps/JSON, remove obvious Whisper loops, and mark uncertain spans as [unclear] rather than inventing words.

Fast path

Run from this skill directory:

cd /Users/mitsuhiko/Development/agent-stuff/skills/audio-transcription
./transcribe-audio.py "/path/to/audio.m4a" --language en --quality balanced

The script:

• stages a stable copy of the input under /private/tmp/audio-transcription-inputs/
• ensures the selected model is cached (downloads only if missing)
• writes txt, srt, vtt, tsv, and json to /private/tmp/audio-transcriptions/<name>-<timestamp>/
• detects obvious hallucination loops and, in balanced mode, reruns with the full model if needed

Useful variants:

# Quick draft, fastest cached model
./transcribe-audio.py audio.m4a --language en --quality fast

# Bad/important audio, slower full model
./transcribe-audio.py audio.m4a --language en --quality best \
  --prompt "Armin Ronacher dictating about AI, data centers, Vienna, Donauinsel, shareholder value."

# Auto language detection when language is genuinely unknown
./transcribe-audio.py audio.m4a --language auto --quality balanced

Cached models

Default model IDs:

• Fast/balanced: mlx-community/whisper-large-v3-turbo
• Best fallback: mlx-community/whisper-large-v3-mlx

Pre-cache / refresh both models:

cd /Users/mitsuhiko/Development/agent-stuff/skills/audio-transcription
./precache-models.py

Verify cache manually:

find ~/.cache/huggingface/hub -maxdepth 1 -type d -name 'models--mlx-community--whisper-large-v3*' -print

If a model is already cached, mlx_whisper should say Fetching 4 files: 100% almost instantly.

Manual command template

If the helper script is not suitable, use this command directly:

mkdir -p /private/tmp/audio-transcriptions/manual
uvx --from mlx-whisper mlx_whisper "/stable/copy/of/audio.m4a" \
  --model mlx-community/whisper-large-v3-turbo \
  --language en \
  --condition-on-previous-text False \
  --word-timestamps True \
  --hallucination-silence-threshold 2 \
  --output-format all \
  --output-dir /private/tmp/audio-transcriptions/manual \
  --output-name transcript \
  --verbose False

For especially rough audio, replace the model with mlx-community/whisper-large-v3-mlx.

Quality checks

Inspect the generated .txt first, then the .srt/.json around suspicious areas.

Red flags that require rerun or cleanup:

• repeated phrases for many lines (eg. in nature loops)
• many zero-duration segments
• avg_logprob is NaN or compression ratios are very high in JSON
• text contradicts obvious context words supplied in the prompt

When finalizing, lightly punctuate and paragraph the transcript, but do not over-edit uncertain content.