does enhancement model has ONNX format ?

[Unmesh28]

Looking for faster inference. Does the model has ONNX runtime ?
any other ways to increase inference speed ??

[vizagite]

distilling mossformer-2 pretrained model to simpler model or quantizing it or maybe making it work towards 16Khz (retrain or can we change architecture surgically idk)

[Unmesh28]

@vizagite I found this :

https://modelscope.cn/models/iic/speech_dfsmn_ans_psm_48k_causal/summary

[vizagite]

DFSMN is very old model, maybe it will be faster, but it cannot achieve the quality of mossformer or such recent models afaik.
there is newer https://modelscope.cn/models/iic/speech_zipenhancer_ans_multiloss_16k_base from alibaba too,

the released pretrained one is small version but still would be on par with this model, not sure of time comparisons, maybe you can test and compare these two for same audio and report back for us.