Noisy audio generated during inference with the provided pretrained checkpoint and config

[zmx-111]

Hi, thank you for releasing this excellent work and the pretrained model.

I am trying to reproduce the inference results using the pretrained checkpoint provided in the repository. I used the inference.py script together with the provided config parameters and the checkpoint checkpoint-epoch140.pth. The model can be loaded successfully and inference runs without errors.

However, the generated super-resolved audio contains a large a

mount of noise, especially in the high-frequency region.