Welcome to ANMF Voice Conversion Demo.
We demonstrate our results of Dictionary Update in an Auto-encoding NMF for Voice Conversion, an attempt to join Deep Neural Networks with Non-negative Matrix Factorization
Dictionary update is a headacke in NMF with high dimensional and huge amount of data. Without dictionary update, a common practice is to (randomly) choose a manageable size of dictionary (say 3000 bases) and skip update procedures. While the latter option, exemplar-based NMF (ENMF) is simple and fast to deploy, it sacrifices voice quality. Our Auto-encoding NMF (ANMF) is a reformulation of the ENMF using an autoencoder. Under this reformulation, we can easily update the dictionary using mini-batch Stochastic Gradient Descent. As a result, ANMF offers better voice quality.
Source
Target
ENMF w/ 3000 basis
Our method
Source
Target
ENMF w/ 3000 basis
Our method
Source
Target
ENMF w/ 3000 basis
Our method