A Modular Approach to Music Generation:
Adding music controls to neural audio compression models
Approach Diagram
Transformation Training
Randomly select new pitch of same instrument during training
Hierarchichal Token Usage
We notice significant increase in quality when more than one hierarchichal token is used doing transformation. Using 2-3 tokens seems to result in about the same quality
Time Difference
We show that transformation from the compresssed audio space is significantly faster than transformation from traditional audio samples
Audio Samples
3 different audio samples tranformed across an octave