Whisper-medium fine-tuned for Teochew ASR
panlr/whisper-finetune-teochew is a fine-tuned whisper-medium checkpoint for Teochew (潮州话 / 潮汕话) automatic speech recognition. The goal is “proper writing” transcription of spoken Teochew — not translation into Mandarin. It’s trained on the open panlr/teochew_wild dataset (about 18.9 hours / 12,500 clips) covering multiple speakers and regional accents.
Two details make this one worth bookmarking if you work with regional language ASR: (1) the dataset uses a custom orthography (“歹看正字法”) to reduce ambiguity that can happen with homophones and variant characters, and (2) the README includes measured character error rate (CER) numbers after fine-tuning (roughly ~12% on validation/test splits). If you want to try it quickly, the model works with the standard transformers ASR pipeline — point it at a .wav file and inspect the returned text.
There’s also a hosted Spaces demo linked from the README, which is handy for a first sanity-check before you wire it into a batch transcription script. If you’ve struggled to find public ASR resources for Teochew (or you need a starting point for another Min dialect), this is a solid “baseline + data + recipe” bundle.
Quick stats from the listing feed: pipeline: automatic-speech-recognition · 10 likes · 32 downloads.
Source listing: https://huggingface.co/models?sort=modified