Task 036: Whisper STT Infrastructure for EP05 Subtitles

Status: DONE

Motivation

EP05 (sm45987761) was uploaded to Niconico on 2026-02-23 but has no subtitles available (Niconico subtitles: {}). YouTube upload is still pending, so no VTT auto-subs either. This blocks both Task 009 (EP05 dialogue attribution) and Task 023 (EP05 full analysis).

The human directive in ideas/ocr_speech_to_text.md calls for building OCR/STT infrastructure as additional subtitle sources. Whisper is the recommended approach for VOICEROID content, which YouTube's ASR handles poorly.

Progress

• [x] Download EP05 audio from Niconico → raw_data/audio/ep05_sm45987761.wav (298 MB)
• [x] Build Whisper JSON parser (whisper.ts) with quality filtering
• [x] Write 19 tests (whisper.test.ts) — all pass
• [x] Create CLI script (run-whisper.ts, npm run whisper)
• [x] Add "whisper" source type to subtitle-types.ts + dialogue-extraction-types.ts
• [x] Codex review: 5 findings fixed (execFileSync, minAvgLogProb, segment sorting, model default, word_timestamps)
• [x] Run Whisper medium model on EP05 audio (~1.42GB model, ~82 min CPU-only transcription)
• [x] Run dialogue extraction (Phase 1) on Whisper output

Results

• Whisper transcription: 341 segments, 325 reliable (16 filtered by quality thresholds)
• Quality stats: avg_logprob -0.212 (excellent), avg_no_speech_prob 0.083 (low), language detected: ja
• Dialogue extraction: 164 merged lines from 325 subtitle entries
• Output files (in raw_data/whisper/, gitignored):

- sm45987761_whisper.json (raw Whisper output, 949 KB)

- sm45987761_subtitle.json (RawSubtitleFile for pipeline)

- sm45987761_quality.json (quality report)

- ep05_lines.json (Phase 1 extracted dialogue, 164 lines)

• Test suite: 704 tests passing (19 new whisper tests)

Scope

1. Download EP05 audio from Niconico (sm45987761) using yt-dlp
2. Run OpenAI Whisper (local, medium model) on the audio to generate Japanese transcription with timestamps
3. Build a Whisper-to-subtitle converter that outputs the same format as VTT parsing (compatible with existing subtitle pipeline)
4. Run dialogue extraction (Phase 1) on Whisper output to produce ep05_lines.json
5. Write tests for the Whisper subtitle conversion

EP05 Metadata (from Niconico)

• Title: SOLAR LINE Part5 END（ソーラーライン）【良いソフトウェアトーク劇場】
• Duration: 1631 seconds (27:11)
• Tags: 最終回, 東北きりたん, SF, 宇宙, ソーラーライン
• Description: 小型船貨物船を操るきりたんが，太陽系を駆け抜けます．
• Comments: 827, Likes: 538, Views: 2614
• Upload date: 2026-02-23

Non-Goals

• Phase 2 dialogue attribution (requires manual review)
• Full EP05 report update (separate follow-up task)
• OCR infrastructure (separate future task)

Depends on

• Task 004 (subtitle pipeline) — DONE
• Task 009 (dialogue attribution) — IN PROGRESS, blocked on EP05 subtitles
• ideas/ocr_speech_to_text.md (human directive)