Whisper-small fine-tuned on Asante Twi (TwiBench)
Decoder-only fine-tune of openai/whisper-small on ~2.5 h of Asante Twi
(Ashesi Financial Inclusion Speech Dataset, CC-BY-4.0), evaluated on a speaker- and
sentence-disjoint test set.
| Test set | WER before | WER after | CER after |
|---|---|---|---|
| Asante Twi (234 clips) | 173.3% | 79.9% | 54.5% |
Twi is not a Whisper language: generate with language="yoruba" as a proxy token,
as used in training.