Ahmad ZushanAI · Systems · Platform

Home/2026 builds/№ 36

36 · Voice & language

Bol Trainer & training kit — hunting the words Whisper gets wrong

Instead of transcribing hours of audio blind, the trainer picks sentences around the sounds the model confuses, scores each word as correct, uncertain, wrong or missing by edit-distance alignment on word timings, and keeps a protected test split. The kit turns the exports into a fine-tuned Whisper, a CTranslate2 export and the context corrector the apps use.

PeriodSep 2026StatusToolingAreaVoice & language

Highlights

  1. Word-level alignment with join and split handling and spelling-variant keys; missed-word spans estimated from the speaking rate of the neighbours.
  2. Exports in Hugging Face audiofolder, NeMo, Kaldi/ESPnet, LJSpeech and TextGrid formats with a dataset card and checksums; 48 kHz takes with quality and speaker metadata.
  3. Training: LoRA rank 32 on Whisper large-v3-turbo with timestamp-aware samples, evaluation with hallucination guards, merge and CTranslate2 export; 60 tests on the trainer and 26 on the kit, the latter hardened after an adversarial review of 17 findings.