Is there anybody uses custom data using librosa for traning LAS

[qute012]

Hi. I want to change dataset and recognize in korean.
I use librosa mel-spectro as extraction features. It shows below.

I didn't modify model parameters and train parameters. But model predicts same class almost one class when i gave some samples.

Below, When i gave sample by using torch.randn, results

It was almost just 2 classes. Is there problems at preprocessing audio file(e.g. mfcc, mel spectro) or models?

anybody knows?