SenseVoiceSmall Hakka
SenseVoiceSmall fine-tuned for Hakka dialect speech recognition.
Architecture
- Encoder: SenseVoiceEncoderSmall (50 SANM layers + 20 temporal pooling layers)
- Frontend: 80-dim fbank, LFR (m=7, n=6), CMVN
- Tokenizer: SentencePiece BPE (Chinese/Japanese/Cantonese/English/Korean)
Original Author
License
Please refer to the original model's license on ModelScope.