andong-li-speech / eabnet

This is the repo of the manuscript "Embedding and Beamforming: All-Neural Causal Beamformer for Multichannel Speech Enhancement", which was submitted to ICASSP2022.

EaBNet

This is the repo of the manuscript "Embedding and Beamforming: All-Neural Causal Beamformer for Multichannel Speech Enhancement", which is to be submitted to ICASSP2022. The network code has been released and you can train it with your own pipeline. The demo website can be viewed at here.

I have rearranged the network code, so if you find any mistakes during the training, please feel free to tell me!

eabnet's Issues

How to convert to waveform from the generated stft

Thanks for your code.
I wonder how to convert from enhanced stft generated from the generator to the waveform so that I can hear the enhanced audio
I know that I need to use torch.istft, but since you conduct power compression to the output of torch.stft, I don't know how to change the enhanced output before applying torch.istft.
I hope you can help me!
Thank you!

多通道输出的标签问题

您好，看了您的论文《EMBEDDING AND BEAMFORMING: ALL-NEURAL CAUSAL BEAMFORMER FOR MULTICHANNEL SPEECH ENHANCEMENT》和对应的代码，有一个疑问：多通道的target是什么？看论文和代码都没有具体说，您是以某一个通道作为整体的target还是多个通道有多个target，然后分别进行处理的？

Data Mixing

I am amazed by your results!

Actually, after finding out that the pre-trained models were not in the directory, I tried to train the models myself.
But, the data mixing stages are a bit confusing for me.

1. Did you locate the microphone array with parallel to room width or length? or with definitely random angle?
2. Did you locate the sources with 0 to 180 degrees from the center of microphone array? or 0 to 360 degrees (so that the target source and the interference can be placed on the opposite side of the microphone array.)?

Could you please answer this question or share the reference code for mixing the data?

Hope you have a great day, bye!

RIR参数问题

你好，我在处理您论文中提到的数据，其中RIR参数中阵列位置没有指明，请问您的阵列的location在哪里呢？
此外，在我把阵列位置设定为房间正中间后，仿真得到的RIR参数后，干净语音clean和rir1卷积，噪声信号noise和rir2卷积，并得到多通道信号，相加之后得到含噪信号noisy，请问这个noisy的SDR是否与SNR接近，
文章table2中：Noisy - - - - - 1.45/29.62/-4.89 1.62/37.71/-1.93 1.74/41.85/0.05 1.87/49.42/2.05 1.67/39.65/-1.18。分别接近（-5dB,-2dB,0dB,2dB）
但是我的仿真结果SDR与SNR在高信噪比时是不接近的,并且stoi也比较大

8k采样率的配置

您好，您能给一下EaBNet在8k采样率下的配置吗？
我修改了这里的sr=8000和fft_num=160，结果会报错
https://github.com/Andong-Li-speech/EaBNet/blob/719f764bbc5128df8cf5eaf59d91e7a2a577939d/EaBNet.py#L768C1-L772C60