Task 285: Whisper Turbo Transcription Upgrade

Objective

Upgrade Whisper transcription from medium to turbo (large-v3-turbo) model for all 5 episodes to improve ASR accuracy. Task 284 built the accuracy measurement infrastructure (Levenshtein-based comparison); EP01 medium achieves 82.6% corpus accuracy vs official script.

Results

EP01 Accuracy Comparison (vs official script)

Turbo improves corpus accuracy by +8.8pp over medium.

Decision: turbo over large-v3

• large-v3 (3GB): ~24 frames/s on CPU — prohibitively slow (~80 min per episode)
• turbo / large-v3-turbo (1.5GB): ~73-180 frames/s — feasible (~10-26 min per episode)
• turbo accuracy already +8.8pp over medium, sufficient improvement

Files Created/Modified

• reports/data/episodes/ep0{1-5}_lines_whisper_turbo.json — turbo transcription lines
• reports/data/calculations/transcription_accuracy.json — regenerated with turbo data
• ts/src/transcription-accuracy.test.ts — 3 turbo accuracy tests
• ts/src/article-content-validation.test.ts — 2 turbo validation tests
• scripts/run_whisper_large_v3.py — batch Whisper processing script
• scripts/process_large_v3_output.sh — pipeline processing helper
• Raw data (gitignored): raw_data/whisper/turbo_raw/, raw_data/whisper/turbo/

Status: DONE