Final Project presentation for the Audio Signal Analysis course at the École Normale Supérieure (Master MVA)!
This presentation is about the paper "ACVAE-VC: Non-Parallel Voice Conversion With Auxiliary Classifier Variational Autoencoder" by Kameoka et al. The paper presents a VAE architecture for voice conversion which has a classifier attached to it. The classifier's role is to force the VAE's output to resemble the new voice class it is trying to emulate. The error is lowered when the VAE and classification error are bettered. The code was mostly adapted from that of the authors' and takes approximately 2.5 hours to run. The results here were presented, and illustrate the strong performance of the authors' approach.
The paper and original code can be found here: https://arxiv.org/abs/1808.05092