Multi-channel Speech Dereverberation using Denoising-Wavenet

model/dwavenet.py includes a PyTorch implementation of the DNN model proposed in A Wavenet For Speech Denoising.

But the training code is written for multi-channel speech dereverberation, not speech denoising.

For the original training code, refer to the authors' repo. It's the Keras + Theano version.

Requirements

Data

This repo requires following data:

• The RIRs for spherical microphone array

  • It can be generated by SMIR generator.
  • The .mat file named hp.dict_path[f'RIR_{room_create}'] should include RIR_TRAIN and RIR_TEST.
  • The shapes of RIR_TRAIN and RIR_TEST are [No. of microphones $\times$ length of impulse response $\times$ No. of source-microphone positions].

• The regularized inverse of the 0-th order mode strength $b_0^{-1}(kr)$ (bEQf)

  • It is required for monitoring validation input and output (using tensorboard).
  • The .mat file named hp.dict_path[f'bEQf.mat'] should include bEQf.
  • bEQf should have the shape of [No. of frequency bins $\times$ 1].

Packages

• python >= 3.7 (or 3.6 with dataclasses backport)
• numpy
• scipy
• matplotlib
• PyTorch >= 1.0
• tensorboardX >= 1.7
• PySoundFile
• librosa
• tqdm
• torchsummary

Dataset

create_mulchwave.py calculates spherical microphone array recordings from speech sources and the prepared RIRs.

Read docstring of create_mulchwave.py for usage.

Training and Testing

main.py is used to train or test DNNs.

_Hyperparameters.l_target is the target field length in the A Wavenet For Speech Denoising.

Read docstring of main.py for usage.

Evaluation Metrics

Source codes for PESQ, STOI, and fwSegSNR are in matlab_lib directory.

Frequency-domain SegSNR is implemented in audio_utils.py.