mlx youtubescribe.stub

---

id: mlx-youtubescribe name: MLX-YouTubeScribe repo_path: ~/Documents/sw30labs/repos/youtube-scribe-mlx remote_url: https://github.com/sw30labs/youtube-scribe-mlx.git primary_language: Python framework: MLX + Whisper + LangGraph ingested: 2026-04-11 last_commit_date: 2026-04-12 category: local-inference-mlx stacks: [mlx, langgraph, converter]

MLX-YouTubeScribe

Purpose

YouTube video/playlist transcription using local Whisper models. Streamlit GUI + CLI. Processes audio, analyzes characteristics, generates transcripts fully offline.

Architecture

LangGraph pipeline (src/main_langgraph.py):

1. Download: Audio extraction via yt-dlp
2. Analyze: Audio characteristics (sample rate, duration, channels)
3. Diarize: Optional speaker diarization via pyannote.audio
4. Transcribe: Whisper transcription via MLX or transformers
5. Optional: Voxtral backend for enhanced quality

Interfaces

• Streamlit Web UI: Interactive GUI for single videos/playlists
• CLI: Command-line interface for batch processing

Key Technologies

• Framework: LangGraph
• Speech-to-Text: Whisper via MLX or transformers
• Audio Download: yt-dlp
• Speaker Diarization: pyannote.audio
• Alternative Backend: MLX-Voxtral
• Web UI: Streamlit

Dependencies

python-dotenv, langchain, youtube_transcript_api, mlx, mlx-audio, yt-dlp, numpy, scipy, transformers, huggingface-hub, pyyaml, mlx-lm, chromadb, torch, langgraph, pyannote.audio, soundfile, demucs, mlx-voxtral

Tags

python, mlx, apple-silicon, local-inference, stt, audio, langgraph

Cross-references

• audiobook-generator — shared audio pipeline
• screenlens — shared media processing