ttlequals0/MinusPod

MinusPod – Self‑hosted ad‑remover for podcasts

What it does

• Runs as a Docker service you host yourself.
• Pulls podcast RSS feeds (or your own audio files), transcribes each episode locally with faster‑whisper or Parakeet (or via a remote API).
• Sends the transcript to a large language model (Claude, Ollama, OpenRouter, OpenAI‑compatible, etc.) which identifies ad segments.
• Cuts the ads out of the audio with FFmpeg, rebuilds a clean RSS feed and optional chapters, and serves the ad‑free episodes through a web UI or a REST API.
• Learns repeat‑sponsor patterns from your manual corrections so the LLM doesn’t have to be asked for the same ad again.

Key AI‑related components

• Transcription – optional local Whisper‑based models (faster‑whisper, NVIDIA Parakeet, OpenVINO) or remote OpenAI‑compatible services.
• Ad detection – LLM prompted to label ad spans in the transcript; supports a verification pass and a “reviewer” UI for human correction.
• Provider‑agnostic LLM – you can plug in Anthropic, Ollama, OpenRouter, or any OpenAI‑compatible endpoint, with per‑stage model/temperature tuning and encrypted key storage.
• Pattern learning – after you confirm cuts, the system builds a sponsor‑pattern database that can auto‑skip future ads from the same sponsor.

Why it matters

• Podcast listeners often tolerate ads; MinusPod removes them automatically, saving time and bandwidth.
• All processing can stay on‑premises, keeping transcripts and audio private.
• The pipeline showcases a practical LLM use‑case: turning a transcript into actionable editing commands.

Getting started (high‑level)

1. Install Docker (GPU version if you have an NVIDIA card, otherwise use the CPU image).
2. Create a .env file with your LLM API key, a base URL, and a master passphrase for encrypting keys.
3. docker‑compose up -d spins up the service.
4. Open http://localhost:8000/ui/ to add podcast feeds, configure models, and review detected ads.

Documentation & extensibility

• The repo ships extensive docs (pipeline details, audio‑cue detection, community pattern sync, backup & security, etc.).
• You can contribute new ad‑pattern manifests, benchmark new LLMs, or extend the REST API.

License – MIT, so you can run, modify, and redistribute it freely.

---

MinusPod is a genuine software project that combines transcription, LLM‑driven ad detection, and audio processing to give users a self‑hosted, privacy‑preserving podcast ad‑removal solution.