sherpa-onnx-sense-voice-ax650

SenseVoice multilingual ASR model converted to AX650 AXMODEL for on-device inference via sherpa-onnx. Supports Chinese, English, Japanese, Cantonese, and Korean.

Model Details

• Original Model: FunAudioLLM/SenseVoiceSmall
• Architecture: SAN-M (Self-Attention Network with Memory) encoder + CTC decoder
• Task: Multilingual ASR + emotion recognition
• Languages: Chinese (zh), English (en), Japanese (ja), Cantonese (yue), Korean (ko)
• Sample rate: 16000 Hz
• Target Chip: AX650 (NPU3)
• Quantization: U16

Files

File	Description
sensevoice.axmodel	AX650 compiled SenseVoice model
tokens.txt	Token vocabulary file
am.mvn	Acoustic model mean/variance normalization
chn_jpn_yue_eng_ko_spectok.bpe.model	BPE tokenizer model

Usage with sherpa-onnx

./sherpa-onnx-offline \
  --sense-voice-model=sensevoice.axmodel \
  --tokens=tokens.txt \
  --provider=axera \
  audio.wav

Performance

Metric	Value
RTF	TBD

Conversion Details

• Pulsar2 Version: 6.0
• Calibration: MinMax with 10 samples
• Quantization: U16

License

Apache 2.0