byrtimo/end-to-end-speech-enhancement-based-on-discrete-cosine-transform

end-to-end-speech-enhancement-based-on-discrete-cosine-transform's Issues

最后一个decoder没有使用sigmoid函数

volume is too low

Hi,
Thanks for sharing this project. I trained a model with my data, which is quite noisy (between -15dB - 10dB 60 hours).Noise reduced well enough, but also speech volume is quite low. I changed the parameters, and then I got pretty much the same results.
(n_window = 512 , stride = 16, selection = 5104 ). Do you recommend any changes that affect the results? Also, I'm just wondering, in Adam optimization many people use beta1 value as 0.9, is there any reason you set it to 0? Last thing I want to ask how do you calculate selection for different n_window sizes? I calculated like this:
selection = n_window * x - stride
x : could be any number.
Is this correct?
Thank you, have a nice day.

这句报错，还有noisy是指的clean和noise合成之后音频吧

grads = opt.compute_gradients(loss_fn, var_list=[var for var in tf.trainable_variables()])