Podcast RSS Whisper Fetcher

A fully local collection pipeline that extracts transcripts from audio-only podcasts. It works from a single input (RSS feed URL + the newest N episodes): it parses the RSS feed (standard XML compatible with Apple/Spotify), downloads each episode's MP3, converts it to 16kHz mono WAV with ffmpeg, and transcribes it on-device with whisper.cpp. On Apple M-series chips, thanks to Metal GPU acceleration, it runs at ~40x real time (43x verified on M2).

When to use: When you want to batch-transcribe a podcast's latest episodes, mine patterns from founder/startup podcasts, archive academic lecture or conference recordings, or transcribe Turkish podcasts (--language tr + ggml-medium).

Output: For each episode, a markdown transcript file with frontmatter (transcripts/<episode_id>__<title>.md) and a _whisper-log.jsonl audit record tracking the collection process. This showcase presents the row-by-row result of a harvest run (episode metadata + transcript summary + performance/status fields) as a single dataset.

Advantage: Quota-free — no OpenAI or Apify API key required, zero API cost. The only cost is disk space (model + temporary WAV ~30MB/episode). With the resume feature it skips previously processed episodes; it automatically discards trailer/intro episodes shorter than 10 minutes.