seorim0 / nunet-tls Goto Github PK

Hello, as a newcomer who has just dabbled in speech enhancement, I really want to read your paper on this program. Excuse me for being stupid, I didn't find your paper on the Internet. Can you share it here?

Firstly,thanks your work very much, I'm study the TFA module recently. I have some problems in the script. 1. the shape of ZF is [1 D] where the script is [T D] 2. the time_seq is the number that add before input x, so why is this done in real time and how the value (32 in script) of time_seq should be set and what it relates to? Thank you again!

Hello,同学你好！

In the last few days, I was trying to change the dataloader from loading .npy huge file to load wav files from folders directly on your DCCRN project. (https://github.com/seorim0/DCCRN-with-various-loss-functions)

Fortunately, I saw a same requset from the Issues (seorim0/DCCRN-with-various-loss-functions#4), which helped me a lot.
Now, I find a function in NUNet-TLS --> tools.py --> minMaxNorm(wav, eps=1e-8), However, is the function result value correct?

tools.py line72-76:
def minMaxNorm(wav, eps=1e-8):
max = np.max(abs(wav))
min = np.min(abs(wav))
wav = (wav - min) / (max - min + eps)
return wav

the wav data should have negtive values, so... should we change the code to like this?

def minMaxNorm(wav, eps=1e-8):
max = np.max(wav)
min = np.min(wav)
wav = (wav - min) / (max - min + eps)
return wav

I mean, should we abandon the abs()?
thank you!!!