Grosbeak

October 16, 2018 ยท View on GitHub

Audio quality manipulation experiment using deep learning.

Goal

Be able to infer high quality, studio level audio recording out of normal recording from normal phone for any type source (guitar, piano, vocal). (Ambitious right? I know)

How to use?

This project requires you to have:

  1. Dependencies listed in Pipfile, use pipenv for convenience.
  2. A bunch of audio dataset to train on, put it in raw_dataset directory.
  3. (optional) Use pycharm if possible

To run this code do the following:

  1. Generate the dataset using generate_dataset.py, run the DEFINE VAR then GENERATE DATASET cell if using pycharm, delete the last CLEAN DATASET part, then run the file otherwise.
  2. Generate the model using the train_model.py
  3. Put audio file in inference_input, then run infer.py to generate output from trained model.