Add FunASR/SenseVoice models to supported model list

[LauraGPT]

Hi! Great C++ runtime hub for multilingual ASR models.

Would you consider adding FunASR models (Paraformer, SenseVoice) to the supported model list?

Available ONNX/GGML models

• Paraformer-large — High-accuracy non-autoregressive ASR, ONNX available
• SenseVoice-Small — Ultra-fast (~70ms/10s), 50+ languages, ONNX available
• Paraformer-streaming — Real-time streaming ASR

ONNX model links

• Paraformer: https://huggingface.co/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-onnx
• SenseVoice: https://huggingface.co/FunAudioLLM/SenseVoiceSmall
• FunASR ONNX runtime: https://github.com/modelscope/FunASR/tree/main/runtime/onnxruntime

Why add these?

• State-of-the-art accuracy on Chinese/CJK benchmarks
• Ultra-low latency (SenseVoice-Small: ~70ms for 10s audio)
• Built-in punctuation and inverse text normalization
• Apache 2.0 license
• 16K+ GitHub stars, widely adopted

References

• FunASR: https://github.com/modelscope/FunASR
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8K+ stars)

[CKwasd]

Hi LauraGPT, thanks for the suggestion.

CrispASR already has FunASR Paraformer and SenseVoice support integrated in the ASR backends list:
https://github.com/CrispStrobe/CrispASR#asr-backends

One small correction on licensing to avoid misleading users: as you clarified in your own reply, the FunASR framework code may be MIT/Apache-2.0, but the FunASR/FunAudioLLM model weights, including Paraformer and SenseVoice, are under the FunASR Model License, not Apache-2.0.

Reference:
FunAudioLLM/Fun-ASR#99 (comment)

The model weights are commercially usable with attribution, but the model license should be listed separately from the framework code license to avoid confusion.

Thanks again for raising this.