fgnt / nn-gev Goto Github PK

The code for the MVDR beamformer requires the atf input, how is this estimated in practice? Is it possible to estimate it also from the signal mask or does it need a DOA algorithm?

what is the "sensor" standing for in the code "nn-gev/beamform.py"

I have some problems when running your CHiME3_simulate_data_patched.m. One of the problems is "undefined function : json2mat ". Can you provide thoses files to us?

Thank you very much for your contribution in the field of multi-channel speech enhancement, which has greatly helped many researchers.

Hi, do you have any plans to use the CHiME-5 dataset to re-train the model and evaluate on CHiME-5?

The string 'dt05_' is suppose to the variable set no? Or the output wav of another set of 'et05_' would be output to the 'dt05_' folder.

Is there a reference to show what the phase correction is doing?

I was having issues running the simulate data code in R2017b since wavread and wavwrite have been removed. I submitted a pull request to address this: #3 if you are interested in using this code with modern MATLAB in the future.

hi . i want to know some about matlab eg : % Compute the STFT (short window) stft_multi();and
estimate_ir() ;apply_ir()......! Can you provide these functions？ i don't pass this matlab code in my data. or Can you provide these functions in python ?

hi,I notice that mvdr in your nn_gev code. I want to compare the performance of mvdr and gev. In beamforming.py ,the function of get_mvdr_vector. How can i set the atf_vector? Thank you for your answer.

Hi, to generate the required data using the MatLab script, it seems there should be a 'data' directory under 'CHiME3' directory that contains the .json files? (but there is not...)

Hi, as you mentioned in the paper, it's also possible to train on clean speech only. I checked the source code but couldn't find any relevant code about this. Could you point me where are you doing that? Or what's the way of doing that if the code is not included? Thank you very much!

Hi,

Are the published pre-trained models (best.nnet) trained with noise-aware data? In other words, is it trained on the simulated data? or the "clean" speech only data?

Thanks in advance.