whisper-small-pld-tsg

openai/whisper-small finetuned on sapinsapin/pld.

Extended run: trained to convergence on the tsg portion of PLD read speech, selected on held-out CER. WER/CER are lowercased on the held-out split; CER is the model-selection metric because Philippine-language orthography varies at the word level. The split is random over utterances, so speakers and prompts overlap between train and test: these are in-domain numbers. Research and non-commercial use only. PLD is licensed CC-BY-NC 4.0 and obtained under a pledge of exclusively research and academic use (Cajote et al., SIGUL 2024, section 3.4); a model trained on it carries those terms regardless of the Apache-licensed Whisper base.

metric	value
cer	0.0477
wer	0.1152

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.