transcribe-offline-meeting

name	transcribe-offline-meeting
description	Processes offline audio recordings from an external mic into the daily log pipeline.
metadata	{"version":1,"tags":["audio","journal","transcription"]}

Transcribe Offline Meeting

Purpose

This skill defines the process for an agent to take a raw .wav or .mp3 file (typically recorded offline via a DJI Mic 2 or Dictaphone), transcribe it using the native ARM64 MLX Whisper path, and then use the AI's natural language understanding to provide a speaker-labeled summary in the LifeRepo DAILY_SUMMARY_WORKFLOW.

Script Usage

The core transcription utility is located at: georgeskills/scripts/transcription/mlx_transcriber.py

PREREQUISITES

1. Configure the durable native ARM64 runtime and model cache in ~/.config/georgeskills/mlx-whisper.env.

Execution

Use the MLX path for transcription. Do not choose WhisperX, whisper-ctranslate2, or generic CPU Whisper as a normal fallback for this workflow. If MLX fails, fix the MLX environment first: native ARM64 venv, working mlx / mlx_whisper imports, and a valid Hugging Face cache.

georgeskills/scripts/transcription/run_mlx_transcriber.sh /path/to/file.wav --outdir /georgerepo/journal/audio/transcripts/

Current defaults for the MLX path:

• model: mlx-community/whisper-large-v3-turbo
• preprocessing: built-in silence trimming / energy-based VAD before transcription
• fallback: if the turbo model fails to load or run, the script retries mlx-community/whisper-large-v3-mlx-4bit
• environment: use the native ARM64 transcription venv. If the default Hugging Face cache is stale or points at a missing volume, set HF_HOME to a stable local cache before running the MLX script.

Bootstrap or repair the runtime with georgeskills/scripts/transcription/bootstrap_mlx_runtime.py. Put the runtime outside disposable repo artifacts and put the model cache on durable storage. Do not use an x86_64 uv-managed Python on Apple Silicon.

Do not use the old CPU transcriber to work around a local setup issue. Treat MLX import/cache/model failures as environment blockers and record the blocker in the daily summary or command notes.

Storage rule:

• Keep Workspace/dji-audio/ raw-only.
• Put generated transcripts under <private-repo>/journal/audio/transcripts/YYYY/MM/.
• The MLX transcriber now defaults to <private-repo>/journal/audio/transcripts/ and auto-redirects the legacy dji-audio/transcripts/ and journal/inbox/ targets back into the canonical journal audio tree.

AI Processing & Export

After the script generates the raw transcript:

1. Read the generated markdown transcript file.
2. Label Speakers: Since machine diarization is often skipped in fast mode, the agent MUST manually identify the speakers (e.g., "George", "Teresa") based on the context.
3. Rewrite & Synthesize:
   • Rewrite the raw transcript into a clean dialogue format if requested.
   • Summarize high-level topics, constraints, and decisions.
4. Format: Output the synthesis under the ## Conversation Milestones header using the personality guidelines in georgerepo/journal/PRIVATE-journal.md.
5. Append: Merge this new block directly into today's Summary.md.