Task 293: OCR Transcription Accuracy Comparison

Status: DONE

Goal

Integrate OCR subtitle text into the transcription accuracy comparison pipeline.

Currently, transcription accuracy is measured for VTT, Whisper-medium, and Whisper-turbo

against the EP01 official script. OCR data (Tesseract 5.3) exists for all 5 episodes

but has not been compared for accuracy.

Completed

1. Added ocrToEpisodeLines() conversion function in transcription-accuracy.ts

- Converts OCR frame-level data to EpisodeLines format

- Skips frames with null/empty subtitleText

- Assigns sequential line IDs (ep01-ocr-001, etc.)

1. Added "video-ocr" to source type union in dialogue-extraction-types.ts and report-types.ts
2. Updated transcription-accuracy-report.ts to include OCR data in comparisons
3. Regenerated transcription_accuracy.json with OCR results
4. OCR accuracy automatically displayed on transcription page (pipeline already handles it)
5. Added 11 new tests:

- 7 unit tests for ocrToEpisodeLines (format, IDs, null handling, timestamps, episode mapping)

- 4 integration tests for EP01 OCR real data (line count, sourceType, accuracy range, comparison)

1. Stats refresh: 292→293 tasks, 2176→2187 TS tests, 2767→2778 total

Results

• EP01 OCR corpus accuracy: 10.1% (21 OCR frames vs 229 script lines)
• Comparison: VTT 68.3%, Whisper-medium 82.6%, Whisper-turbo 91.4%, OCR 10.1%
• Tesseract OCR on Japanese anime subtitles is extremely poor — establishes baseline for improvement