Rong Xiaobin
Rong Xiaobin
Hi, your operation looks correct. It might be normal for the voice to be over-suppressed when the SNR is low, as this is a common issue with DNN-based SE. By...
你好，目前的计划是进一步优化该模型在SE中的性能，暂无拓展到AEC的计划~
@TungyuYoung Yes, it's a great direction. I have tried vanilla knowledge distillation but improved little. So you have done it successfully?
@ercandogu-elevear Yes, I tried expanding GTCRN as teacher models. I also found similar phenomena of bad convergence when expanding GTCRN to a certain scale, for example, 6 GMacs. Actually, the...
These codes were written quite a long time ago, so I have forgotten the details. However, when I tried running your code, I encountered the following error: `TypeError: StreamConvTranspose2d.forward() missing...
@qianjr2002 Thank you so much for your contribution! I do apologize if this may seem a bit inconvenient, but this repository does not need training codes as they are available...
32窗长和16步长分别指32ms和16ms，对应512和256点
对 offline 的情况，音频时长 9s，表中的时间不是 RTF，仅是推理时间。RTF 是 推理时间/音频长度。 对 online 的情况，是逐帧处理的，因此音频长度为一帧（16 ms），RTF也是 推理时间/音频长度。