Fine-tuning Whisper for Danish phone calls: a LoRA post-mortem (2026)
Fine-tuned whisper-large-v3 on CoRal-project/coral-v3 (Danish speech, telephone-codec augmented) with LoRA. Training loss got stuck at ~35-40 for three attempts running — looked like large-v3 itself was unstable under LoRA. It wasn't: the warmup schedule was too long relative to how far training actually ran. Fixed, the model nearly halved WER on held-out Danish speech versus the untrained base. Then a systematic 30-call audit on real phone audio found a real fabrication rate the benchmark number didn't show — published anyway, with an honest model card.