search

Qingzheng-Wang / Dual-Window-SE

An implement of STFT-Domain Neural Speech Enhancement with Very Low Algorithmic Latency of Zhong-Qiu Wang et al.
MIT License
12 stars 1 forks source link

Is there an error in the code about loss in train.py? #1

Open Hawkkkkkk opened 12 months ago

Hawkkkkkk commented 12 months ago

Excuse me, I would like to ask why both loss1 and loss2 in train.py are using l_ri_mag? Shouldn't loss2 use l_wav_mag instead?

Qingzheng-Wang commented 12 months ago

I have to confess that it is just an unfinished implementation. I tried to implement the original paper under the single-channel situation. The original paper is based on multi-channel, but I didn't have a relevant dataset. So it is just an imperfect reimplementation. I did not strictly follow the original paper. I guess the only valuable part of this repo is the model.py, which is according to the paper. Due to the lack of the dataset, and I just want to focus on single-channel, I will not complete it in the future. Sorry. Back to your question, I guess you are right! It is my fault.