whisper-medium-yecs-baseline

openai/whisper-medium fine-tuned on the YECS Yoruba–English code-switched speech corpus (frozen training split, 75999 utterances; internal dev held out).

Training

• Language token per utterance: <|en|> or <|yo|> by majority of word tags (tie → en); decode with language auto.
• lr 1e-05, warmup 500, batch 16 × accum 2, bf16, ≤ 10 epochs, early stopping (patience 3) on greedy dev WER.
• Selected step 3000 of 6000 (2.53 epochs). Seed 42.

Internal dev results (beam 5, language auto, %)

wer	cer	di_wer	err_yo	err_en	pier_yo	switch_near_err	switch_away_err	ins_per_word
13.64	4.60	10.00	17.08	7.83	18.91	12.04	12.40	1.40

Scoring: tone- and switch-aware protocol (NFC, punctuation removed, diacritics kept).

Data licence

Trained on YECS, released under NOODL-1.0. Check its terms before use.