fjiang9 / nkf-aec Goto Github PK

您好

我使用您的checkpoint在ICASSP 2021 AEC Challenge blind_test_set clean上验证，没有取得相同的得分

ICASSP 2021 AEC Challenge blind_test_set clean:
https://github.com/microsoft/AEC-Challenge/tree/20274d3a85f752d3fa41ed8e8915e73a5e39b7ff/datasets/blind_test_set/clean

AECMOS:
https://github.com/microsoft/AEC-Challenge/blob/main/AECMOS/AECMOS_local/Run_1663915512_Stage_0.onnx

echo score: 3.32 degradation score: 3.10

Hi,

I am trying to implement the training process described in your arxiv paper. I wonder if you can kindly provide some more details about your training setting:

1. I note the farend talk are randomly sampled to be 1 second, and the nearend talk between 0.5 - 1 second. When you mixed them to produce double talk, did you randomly select a position in the echo to insert the nearend talk?

2. What optimizer did you use? What was the setting of other hyper-parameters of the optimizer apart from learning rate (gradient clipping, momentum, etc)?

3. The echo generated through full convolution is longer than the farend signal, as the length is farend length + rir length - 1. Did you do any clipping on the echo, or you just retain the raw convolutional result?

4. Did you do any random scaling for signals in the synthetic fold of the AEC challenge, before sampling them to create nearend and farend talks?

Thank you very much!

尽管该算法被作者命名为NKF，并且可能作者从kalman公式得到启发进行算法设计。但是从最终迭代方法上看，其行为也可以看做
在通过网络去改变自适应滤波器的步长，而网络输入是可能与步长值相关的量，从这个角度看，网络输入有更多可能，且算法似乎与kalman并无联系。

Hi there, thank you for providing such a good project . I tested your pre-train model and found it work well in my real recording data, and I was wondering how many hours data did you used in train your model , thanks!

Hello, I have replicated your work but get bad result. Could you describe your training process? I modified the test code for training and found that the results were poor.

Hello! The work looks very promising, thanks for research. Did you try to extend the network for fullband audio (48 kHz)? Or this architecture is not suitable for such wide datastream?

'w' will diverge during training, resulting in an INF situation. The loss will be 'nan' or a very big number. How to solve it? I try to add a 'Tanh' after W, which is useful.

Hi,

Thank you for your great project. I have a question about the frequency-domain forward pass in your paper/code. I note your model simply multiplies stft of input signal and filter weights, and the result is an approximation of time-domain convolution. I wonder if you have tried any exact frequency-domain forward pass models, such as overlap saving? And if yes, what's its effect on the performance?

Thanks

我发现在训练过程中，总是存在一些数据会在迭代过程中发散。并且，在训练过程中，loss总是根据数据的不同，来回震荡。请教下作者针对这两个问题，有什么解决办法呢？

In the paper, you have said "The duration of the far-end signal is randomly clipped to 1 s, while the near-end signal is randomly segmented to 0.5�~1 s."
What's the point of doing this?

你好，请问我用自己的中文对话音频测试，发现回声没有被消除，也手动对齐了，请问可能是什么原因呢，还是对汉语需要自己训练模型呢

❤ (ɔˆз(ˆ⌣ˆc)

Although the number of parameters is small, the cost of NKF is a little bit high, especially as the number of taps increases. Have you completed the model deployment on the actual device? Do you have any ideas for simplifying the model?

Can the training stage directly use this forward code to train the model? What is there to pay attention to? Thanks!

您好，在复现模型训练的过程中，遇到了损失函数值异常的问题，有以下几个问题想请教一下作者：

1. 训练数据直接从AEC-challenge官方下载（我用的链接是https://github.com/microsoft/AEC-Challenge/tree/main/datasets/synthetic_fullband ）完成之后，我进行了将采样频率从48khz降到16khz操作，同时模型在读取训练数据(近端为mic.wav，远端为echo.wav，真值为target.wav)的时候会随机选择1s的片段（后续因为nan的出现改成了随机选择0.5s），除此之外作者还有对数据进行其他的操作吗？您是否也是用这样的方式获取到的训练数据？如果不是的话想问一下您是从哪里下载得到的数据？
2. 请问作者是用什么样的方式对神经网络的参数进行初始化的？在我复现的过程中发现如果不是用全0进行weights和bias参数初始化的话损失函数计算会出现nan的情况，您是否也遇到过这样的情况？请问是怎么解决的？
3. 论文里提到损失函数对echo进行L2损失函数计算，是否可以直接调用nn.MSELoss()进行计算？根据论文来看似乎并没有对回声消除的结果直接进行损失函数计算，那预测值是神经网络的输出还是需要经过其他的操作？真值又是如何获取的呢？
   期待您的解答。

请问作者能否将训练模型公开一下呢，我自己写的训练模型代码存在较大问题，第一个epoch就出现nan的情况解决好久也没处理好

I tried to reproduce your results, but I found that there are two phenomena (bad results) in some real sequences, both of which also appear in the model you provide.

1. may be that the filter diverges (the estimated echo energy is unusually high);
2. the estimated echo have a howling-like sound (which doesnt exist in the reference signal);
   Most of these two happens at the beginning of the sentence(but not necessarily at the beginning of the sequence).

Did you find anything like this?

could you please explain that why did you put the init step "self.kg_net.init_hidden(B * F, device)" in the forward function of class NKF? This will put gru weights to be zero, is it OK when training ?