Turn any topic into long multi-part NotebookLM podcasts — with music — and build a searchable dataset of prompts + transcripts + speaker diarization + sources. Comes with a Claude Code skill.

Not another "NotebookLM alternative". This uses the real NotebookLM: it automates it end to end, and captures everything it produces into a reusable dataset that doesn't exist anywhere else.

The Telegram bot in action (menu-driven):

The pipeline, visually — NotebookLM forges the episodes, music drops in, Claude mixes, Telegram delivers:

The NotebookLM app makes one podcast at a time and throws away the prompt. This pipeline:

• Generates podcasts in parts from a single topic (deep-research → N episodes), each with a custom prompt
• Adds music — intro / stinger between parts / background under the hosts — chosen from your own tracks (pure ffmpeg, no re-transcription)
• Builds a dataset: for every podcast it pairs the real prompt + word-level transcript + who-said-what (2-speaker diarization) + source links & extracted markdown. This prompt→transcript dataset is not available online — you build your own.
• Batch everything: download all your Audio Overviews, transcribe them all, recover all prompts — each step resumes where it left off.
• Optional Telegram bot to drive it from your phone with inline buttons.

Nobody combines NotebookLM → podcast → music + diarized dataset. Verified across the whole GitHub NotebookLM landscape.

{
  "title": "How data travels the network",
  "prompt": "Impersonate each step... no intro/outro... topics: socket programming...",
  "transcript": "[SPEAKER_00] Welcome... [SPEAKER_01] Exactly, so...",
  "n_speaker": 2,
  "source_links": ["RFC 791", "Beej's Guide to Network Programming"]
}

On Colab (does everything):

1. Upload storage_state.json (your NotebookLM auth from the desktop CLI) to Drive PodcastLab/
2. Runtime → T4 GPU → run cells 1 → 2 (download) → 3 (transcribe+diarize) → 4 (dataset)

Telegram bot (optional, on your PC):

notebooklm login                 # one-time browser login
pip install python-telegram-bot  # + ffmpeg in PATH
# put TELEGRAM_TOKEN in .env (copy from .env.example)
python bot.py

The underlying notebooklm-py exposes 9 output types — this repo can be extended to generate, from the same topic, not just podcasts but video, quiz, mind-map, flashcards, slides, infographic, report.

• Python 3.10+, ffmpeg
• A Google account with NotebookLM access
• Colab (free T4) for transcription/diarization
• HuggingFace token (free) + accept pyannote/speaker-diarization-3.1 terms for speaker labels

• Uses the unofficial NotebookLM API (via notebooklm-py). Google can change it anytime. For personal / educational use — respect NotebookLM's Terms of Service.
• Auth (storage_state.json) expires ~24h; refresh with notebooklm login on the desktop.
• Language: pipeline defaults to Italian podcasts (language='it'), trivially changed.

MIT