transcribe

name	transcribe
description	Transcribe audio files using the local `transcribe` CLI (WhisperX large-v3 + diarisation + optional LLM correction).
title	/transcribe
parent	Skills
permalink	/skills/transcribe/
nav_order	19

Transcribe Audio

Delegate to the local transcribe CLI from ~/dev/personal/transcribe/ (symlinked into ~/.local/bin/, so callable from anywhere). Do NOT call whisper-cli directly — the CLI uses a better model (WhisperX large-v3), adds speaker diarisation, EBU R128 normalisation, and resumable caching.

Your Task

Given an audio file path as argument:

1. Resolve the file path — expand ~ and verify the file exists.
2. Run transcribe:

   transcribe <audio-file> [lang] [output-dir]
   

   • lang: fr by default, auto to autodetect, or any whisper language code.
   • output-dir: . by default. Pass an explicit directory if the user specifies one or if the audio is somewhere they don't want output files dropped.
3. Report where outputs landed — out/<safe>.txt, .srt, .json, etc. The basename is sanitised (ASCII, lowercase, dashes, ≤60 chars).
4. Offer LLM correction if the transcript looks rough or the user cares about accuracy:

   correct <out/...json> [model]
   

   Default model is mistral-nemo:12b. Produces <safe>.corrected.txt and <safe>.diff next to the JSON.

Guidelines

DO:

• Always use the transcribe CLI, not raw whisper-cli or ffmpeg + whisper-cpp pipelines.
• Trust the CLI's caching — re-running the same command resumes from the last completed stage, so it's safe to re-invoke.
• Surface the output path so the user can open the files.
• Suggest correct as a follow-up step, don't run it automatically (it's slow and uses Ollama).

DON'T:

• Don't reimplement the pipeline inline. If something's broken, fix it in ~/dev/personal/transcribe/.
• Don't pass non-WAV files through manual ffmpeg conversion — the CLI handles all input formats.
• Don't worry about the medium-model whisper-cpp setup — it's been replaced.

Input Handling

• Single argument: audio file path (e.g., /transcribe ~/Downloads/audio.opus)
• With language: file path + language code (e.g., /transcribe ~/Downloads/audio.m4a en)
• With output dir: file + lang + dir (e.g., /transcribe ~/Downloads/audio.m4a fr ~/transcripts/)
• No argument: ask the user for the file path.

Tools Used

• Bash: transcribe and correct CLIs (in ~/.local/bin/).

Dependencies

Set up once per machine — see ~/dev/personal/transcribe/README.md for full instructions:

• uv tool install whisperx
• Hugging Face token at ~/.config/whisperx/token (for diarisation; delete the file to skip diarisation)
• brew install ollama, ollama pull mistral-nemo:12b (only needed for correct)

Example Usage

/transcribe ~/Downloads/voice-message.opus
/transcribe ~/Downloads/meeting.mp3 en
/transcribe ~/Downloads/seance.m4a fr ~/transcripts/