transcribrr

Local audio/video transcription and summarization on Apple Silicon using MLX. Three standalone shell scripts and an interactive MLX model launcher.

Requirements: Apple Silicon Mac, ffmpeg, Python 3

---

Scripts

transcribe.sh — Transcribe audio to text

./transcribe.sh recording.mp3

Uses MLX Whisper (GPU-accelerated on Apple Silicon). Prompts you to select a model size, then writes a transcript with metadata.

Model	Parameters	Notes
tiny	39M	Fastest
base	74M	Fast
small	244M	Recommended
medium	769M	High quality
large-v3	1550M	Best quality
turbo	809M	Near-large, much faster

• Output: <basename>_transcript_<model>.txt
• Progress log: <basename>_transcription_<model>.log
• Dependencies auto-installed into .venv/ (mlx-whisper)

cleanup-transcript.sh — Clean up a raw transcript

./cleanup-transcript.sh recording_transcript_small.txt

Runs the transcript through a local LLM to fix spelling, grammar, and punctuation; removes filler words; adds paragraph breaks at topic transitions. Preserves the speaker's voice.

Model	RAM	Notes
Llama 3.2 1B	~1 GB	Fastest
Llama 3.2 3B	~2 GB	Fast
Llama 3.1 8B 4-bit	~5 GB	Recommended
Llama 3.1 8B 8-bit	~8 GB	Higher quality

• Output: <basename>_cleaned_<model>.txt
• Requires .venv/ with mlx-lm (installed by summarize-transcript.sh)

summarize-transcript.sh — Summarize a transcript

./summarize-transcript.sh transcript.txt
./summarize-transcript.sh --install   # pre-download default model

Summarizes using Qwen 2.5 on Apple Silicon. Handles transcripts of any length via intelligent chunking (single-pass up to ~20k words, multi-pass synthesis beyond that).

Models:

Model	RAM
Qwen 2.5 7B 4-bit	~5 GB
Qwen 2.5 14B 4-bit	~9 GB
Qwen 2.5 32B 4-bit	~20 GB (default)
Qwen 2.5 32B 8-bit	~34 GB

Summary styles: Executive, Detailed, Bullet-only, Chapter, Blog Post (default)

• Output: <basename>_summary_<model>_<style>.md
• Dependencies auto-installed into .venv/ on first run

mlx-chat.sh — Interactive MLX model launcher

./mlx-chat.sh ~/mlx-data

Pass a data directory for the venv and model cache. Lists installed and available MLX models, then runs in terminal chat mode or OpenAI-compatible server mode (for use with VS Code Copilot, etc.).

---

Typical workflow

# 1. Download audio (e.g. with yt-dlp)
yt-dlp -x --audio-format mp3 "https://youtube.com/watch?v=..." -o talk.%(ext)s

# 2. Transcribe
./transcribe.sh talk.mp3

# 3. Clean up the raw transcript
./cleanup-transcript.sh talk_transcript_small.txt

# 4. Summarize
./summarize-transcript.sh talk_transcript_small_cleaned_llama3.1-8b-4bit.txt

All scripts install their Python dependencies into .venv/ on first run — no manual setup required.