Whisper Corruption Recovery

Overview

Whisper on audio ≥2h can enter repetition loops — in-segment phoneme repeats or cross-segment multi-line loops (e.g. A-B-A-B). The whisper-corruption-recovery skill provides a BFS Parallel Divide-and-Conquer (D&C) engine with sub-1-second mathematical base-case guarantees, [?] text prepending, and visual frame context extraction.

Automated Recovery Scripts

1. BFS Parallel Audio Recovery (scripts/fix_hallucinations.py)

Scans transcripts for in-segment phoneme loops and cross-segment multi-sentence loops ($A-A-A-A$, $A-B-A-B$, $A-B-C-A-B-C$), chunks corrupt ranges into 30s tasks matching Whisper's native context window, and executes parallel BFS D&C recovery down to sub-1-second slices.

# Single GPU (Fastest: RX 7600 Vulkan matrix cores)
python3 scripts/fix_hallucinations.py <whisper_json> <audio_wav> --devices 0 -w 3 -o recovered_transcript.json