PastNoise 檢索 206M+ Podcast 字幕行並配對 PubMed 與 Wikipedia 引文
Indexing 206M+ Podcast Subtitle Lines and Pairing Them with PubMed & Wikipedia Citations
作者 nonacosa 建立 PastNoise(pastnoise.com),可搜尋 206M+ 帶時間戳的 Podcast 字幕行,將 5 至 6 位不重複主要講者的片段編排成連續瀏覽頁,並列呈現 Wikipedia、PubMed、Crossref 和 OpenAlex 引文。
When I research a topic through long-form podcasts or university lectures (say, sleep architecture, creatine for cognition, or semiconductor supply chains), I keep hitting two annoying walls:
- Searching YouTube drops you into three-hour episodes full of sponsor reads and detours.
- Dropping a link into an LLM summarizer strips away the speaker's actual tone, caveats, and source provenance.
Over the past few weeks I built PastNoise to scratch that exact itch. Instead of generating another text summary, it searches across 206M+ timestamped spoken subtitle lines and stitches 5 to 6 non-duplicate primary speakers into a single continuous screening room, paired with side-by-side Wikipedia, PubMed, Crossref, and OpenAlex citations.
Three Small Engineering Choices That Made the Biggest Difference
1. Filtering sponsor reads before picking clips (not after)
Early builds looked fine until a flagship topic pulled two Hostinger and Tonal ad reads from top podcasts. Moving a strict sponsor/outro detector ahead of act selection and hydrating 16 candidate windows first (then backfilling clean distinct channels) dropped sponsor contamination to zero across all 58 indexed screenings.
2. Rendering full bilingual transcripts in SSR DOM with #act-X-t-Y citation anchors
Single-page tabs that unmount transcript DOM hide 95% of spoken text from search engines and make it impossible to quote a speaker accurately. Keeping every chapter's timestamped transcript inside lightweight <details> accordions with click-to-copy citation links lets readers jump straight to the exact second and verify the primary quote.
3. Preserving original subtitle sequence IDs (origSeq) when splitting long cues
Auto-captions often pack 220+ characters into one cue. Splitting cues down to readable 96-character lines broke pre-translated subtitle lookups whenever the original sequence index passed 100. Keeping an explicit origSeq field alongside sequential display IDs solved alignment cleanly for English, Simplified Chinese, and Traditional Chinese (via OpenCC at zero extra LLM cost).
Open Assets You Can Reuse (CC BY 4.0)
If you prefer local notes or want clean benchmark data for citation-grounded RAG experiments, I exported the entire 58-topic curated archive under CC BY 4.0:
- Obsidian Knowledge Vault (58 dossiers + 25 Thinker MOCs): github.com/nonacosa/pastnoise-obsidian-vault
- Hugging Face Dataset (491 ad-filtered clips with YouTube timestamps + DOIs): huggingface.co/datasets/pkwenda/pastnoise-primary-citations
- Free Hydration Utility: Galpin Equation Hydration Calculator
- Full Interactive Screening Room (free without login): pastnoise.com
Curious how other builders here handle long-form audio/video retrieval and hallucination checks without losing the original speaker's voice.
來源:DEV Community · dev.to