When converting a model that takes seq_length as an input, make sure to set seq_length to torch.tensor([1]). This is because some layers in the model may use the actual numeric value of seq_length as a parameter, and any value other than 1 cannot be broadcast to arbitrary shapes. Using other numbers will therefore cause model errors during dynamic inference.

Using OpenVINO as backend

Add Readme for OpenVINO

frontend param "dither" must be 0.0 for inference, otherwise the frontend will add random noisy to the input audio