This repository contains a few musical examples generated from an alternative of VQ-CPC that I implemented during my research internship at Sony CSL Paris. Networks have been trained on the NES Music Database.
This repo contains 4 folders:
clusters/
that regroups subsequences from the dataset that are encoded by the same codebaseline/
that contains generations produced by a baseline Transformer without codesVQCPC/
that contains variations generated by my modellinear_VQCPC
that contains variations generated by my model, with the relative attention function replaced by a linear attention
For each variation, the original sample and the variation are concatenated in the same audio file, with a pause of 1 second between the samples.