Corpus
YouTube lectures on the 黃帝內經 and related TCM classics by 徐文兵 (with 梁冬), 游朱義,
南懷瑾 and 倪海廈. Video ids, titles, channels and durations come from the user's pipeline
metadata (talks/*/talk.json, _imports/*/inventory.csv).
Transcription provenance
- 徐文兵, 南懷瑾, 倪海廈: local faster-whisper large-v3 (CUDA, float16), 600 s chunks, converted to Traditional Chinese with OpenCC s2twp. No diarization. Not proofread against audio.
- 游朱義: YouTube author-uploaded or transcript-panel captions (non-generated), zh-Hant.
- 倪海廈: additional automatic page anchors to the printed 《黃帝內經》講課實錄 (string similarity, not forced alignment). The manuscript PDF is not in this repository.
- Pipeline QA (low-confidence flags, gap repairs, before/after audits) is preserved per
talk under
sources/.
Processing
- L1 corrects ASR errors using context, the bundle's review flags, and a cumulative glossary
(
_index/glossary.md, "Seen as" column documents every correction pattern). Uncertain readings are marked[?]; speaker labels are inferred from content only. - L2–L4 are derived solely from L1 and cite timestamps; editorial context is marked 編按.
- Processing was done by LLM agents (Claude, possibly Codex) following
docs/spec/; outputs have not been reviewed by a domain expert unless noted in the file.
Limitations to state in any publication
- Transcripts are unverified machine drafts; quotations should be checked against audio.
- Timestamps are ASR estimates tied to the original upload, not remasters.
- Speaker attribution in dialogues is inferred.
- Speakers' statements are documented, not endorsed; nothing here is medical advice.