Modified and unofficial implementation of the audio super-res network proposed by Kim and Sathe.

This repo is modified and unofficial implementation of the audio super-res network proposed by Kim and Sathe [1]. The original discriminator network has FC, dropout, and leakyReLU layer on top of 8 downsampling blocks. To reduce the network size, I replaced that part with a global maxpooling layer.

• python 3.7.9
• pytorch 1.5.0
• torchaudio 0.5.0
• resampy 0.2.2

1. Download pretrained model (the model is trained for Speaker1 of VCTK dataset (p225) for 200 epochs, and performs upsample 8kHz to 16kHz).

2. Run python 04.extend.py --checkpoint mugan-generator-200.tar --gpu input.wav output.wav to perform superresolution.

Edit model_settings.py to match desired experimental setup.

• supersampling_rate()
• conv_kernel_size(): a list of filter sizes for multiscale conv. layers
• conv_out_channels(): a list of output feature sizes for the corresponding downsampling/upsampling blocks
• out_conv_in_channel(): feature size of the conv. layer on the top of the autoencoders and generator networks.
• out_conv_kernel(): kernel size of the conv. layer on the top of the autoencoders and generator networks.
• the first constant in sample_segment_length(): duration of training/validation sample size in seconds.

Run 01.train-autoencoder.py to train the autoencoder then run 02.train-extender.py to train the GAN.

Simple evaluation and prediction program are also provided. Run 03.eval-extender.py for evaluation and 04.extend.py for prediction (make super resolution of any .wav file)