fbank window function discrepancy between training and deployment

[tomato18463]

The window function for fbank feature here is default setting of torchaudio.compliance.kaldi.fbank, which seems to be povey window. However, the fbank window function in the deployment code here is hamming window. Do you have any empirical results on the influence of this? Will it cause a performance drop? Thanks for your help!

[mlxu995]

Thank you. It seems there is a mistake about the window function in fbank calculation.
I tested the onnx model using the two types of window functions mentioned above. Using the hamming window, there are 21038 files successfully woke up, compared to 21069 files using the provey window (The total number of testing files is 21282).