Thai Wav2Vec2 with CommonVoice V8 (newmm tokenizer) + language model
wannaphong/wav2vec2-large-xlsr-53-th-cv8-newmm is a Thai automatic speech recognition model built by fine-tuning facebook/wav2vec2-large-xlsr-53 on Thai Common Voice data. It’s positioned as a “CV8 refresh” of earlier Thai wav2vec2 work: the author describes re-splitting the V8 dataset, reusing earlier V7 splits, and publishing results for both tokenization schemes (newmm and deepcut).
The model card is worth reading because it includes concrete error rates and shows how much a language model helps for Thai decoding. On the Common Voice V8 test set, the reported WER drops from ~16–17% to ~12.6% (newmm) when adding the LM, and the CV7 benchmark numbers show a similar pattern. If you’re building Thai transcription (customer support calls, media indexing, captions), a good first try is to run decoding both with and without the language model and compare the kinds of mistakes you get—Thai word segmentation choices can move WER a lot even when the underlying audio model is the same.
Quick stats from the listing feed: pipeline: automatic-speech-recognition · 4 likes · 2272 downloads.
Source listing: https://huggingface.co/models?sort=modified