openai/whisper-small finetuned on sapinsapin/pld.

Trained for 2000 steps on 10000 clips (batch 8×2, lr 1e-05, fp16 + gradient checkpointing). WER/CER are on the held-out split, lowercased; CER is the model-selection metric (Taglish orthography varies at the word level).

Trained with finetune_asr.py from the halohalo pipeline; the dataset adapter normalizes each corpus to (audio@16k, text, speaker_id) so corpora are swappable with a --dataset flag.