Deep learning methods for voice processing: Neural vocoding for voice transformation

For some years, the state-of-the-art in speech synthesis and processing has been dominated by data-driven methods and deep neural networks. The use of ever larger amounts of data allows the exploitation of ever more parameters, leading to ever better results. Unfortunately, the increasing computational complexity hinders the widespread application of these models.

In the first part of the talk, we will present our research into data and computationally efficient voice transformation with deep neural networks. We will introduce the Multi-band Excited WaveNet, a deep neural network that integrates a WaveNet into a classical source-filter model. The discussion will motivate model structure and training losses. We will describe the deficiencies of the proposed model and briefly reflect on perspectives considering the rapidly evolving state of the art in neural vocoding.

The second part will then demonstrate ongoing research into applications of the neural vocoder, combining it with dedicated models for intensity, pitch, expressivity or identity transformation.

Bio: Axel Roebel is director of research IRCAM and head of the Analysis/Synthesis team. His research activities center around voice and music synthesis and transformation with strong focus on artistic and industrial applications. After many years or research into various signal processing algorithms he now has shifted his focus towards data driven methods.