a question about the spectrum of sph_est.wav

[Captain2xxx-coder]

Hi,

it's a great job and helps me a lot ! But I have come across a question that the spectrum of estimated speech (sph_est.wav) has stripes shown in the following picture. Besides, my validation set loss didn't change during training and was different from the training set loss. Do you have any ideas to figure out these problems?

Thank you so much!

[JupiterEthan]

Hello, the stripes look like checkerboard artifacts caused by strided convolution. Have you tried to set the kernel size to a multiple of the stride size for all convolution and deconvolution? For example, if you use a stride of (1, 2) for convolution/deconvolution, you can set the kernel size to (1, 4) instead of (1, 3). This will likely alleviate the problem.

Regarding the validaiton loss issue, did you apply any normalization/rescaling to the data? If so, did you make sure that the way of scaling/normalization is the same between training and validation?

[QQQQQQQQY]

Hi, 你好，

it's a great job and helps me a lot !这是一份很棒的工作，对我帮助很大！ But I have come across a question that the spectrum of estimated speech (sph_est.wav) has stripes shown in the following picture. 但我遇到一个问题，估计语音的频谱（sph_est.wav）有如下图所示的条纹。 Besides, my validation set loss didn't change during training and was different from the training set loss. Do you have any ideas to figure out these problems? 此外，我的验证集损失在训练期间没有改变，并且与训练集损失不同。您有什么想法来解决这些问题吗？

Thank you so much!太感谢了！

Hi, is the problem solved that the validation set loss did not change during the training period