sapinsapin/whisper-small-fsc-pld-fil

whisper-small-fsc-pld-fil

`openai/whisper-small` finetuned on `sapinsapin/pld`.

Whisper-small finetuned for Filipino on two corpora together: the Filipino Speech Corpus, which holds studio read, spontaneous and machine-directed speech, and the Filipino portion of the Philippine Language Dataset, which holds prompted read speech. 25,000 clips split equally between them, 5,000 steps, selected on held-out CER.

Why it exists. Each single-corpus model is excellent at home and poor away. Scored on each corpus's held-out test set separately, 300 clips each, CER after lowercasing and removing punctuation and diacritics:

It beats the FSC-only model on FSC and stays within 0.3 points of the PLD-only model on PLD, while removing the cross-corpus drop in both directions. The metrics table above this summary is the training-time evaluation on a blended test set with a looser normaliser, so it reads higher than the per-corpus numbers.

Caveat. Splits are random over utterances, so speakers and prompts overlap between train and test. These are in-domain numbers.

Licence: research and non-commercial use only. Both corpora are University of the Philippines Diliman DSP Laboratory resources distributed for research use, and the Philippine Language Dataset is licensed CC-BY-NC 4.0 under a pledge of exclusively research and academic use (Cajote et al., SIGUL 2024, section 3.4). A model trained on them carries those terms regardless of the Apache-licensed Whisper base.

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.