wqi / bdsr Goto Github PK

View Code? Open in Web Editor NEW

5.0 5.0 1.0 10.48 MB

Audio Bit Depth Super-Resolution with Neural Networks

Python 3.18% Jupyter Notebook 96.82%

bdsr's People

Contributors

Stargazers

Watchers

Forkers

zhanfengdog

bdsr's Issues

Write compatible data loader for VCTK single-speaker dataset

Use wavenet's data set creation tool on our differing bit depth audio

Generate baseline results for Music and VCTK single speaker datasets

Modify wavenet to have different input/output sizes

Extend receptive field to capture information from future audio samples

Our receptive field should look like that from the desnoising WaveNet paper and extend both forward and back.

Write script to evaluate bit-depth upscaling performance

Should compute some sort of quantitative metric like SNR.

Modify wavenet implementation to take in training LR-HR pairs rather than a single audio sample

Generate bit-depth downsampled dataset from VCTK

Kick off initial full-scale BDSR training on Azure

When everything is fully implemented, someone needs to kick off the full training run for BDSR on Azure. We can use a server with more power if necessary.

The first full-scale training run should train on the Mehri piano dataset using the given splits. We will then evaluate it on the associated test set.

Add new loss function for 16 bit depth

Convert output from softmax distribution to real values

Update input and output components of wavenet to work with BDSR

While the data loader is currently working with the Mehri dataset, the input/output dimensions of the network haven't been updated so the forward pass doesn't work yet.

Fixing this might be as simple as just changing the dimensions, but could be significantly more complicated. If it ends up being a really difficult task, just spin out more issues on GitHub and we can all pitch in.