Reproduce “Eigenvoice Synthesis based on Model Editing for Speaker Generation” on OpenResearch

About this paper

Masato Murata, Koichi Miyazaki, Tomoki Koriyama, Tomoki Toda

A novel DNN-based eigenvoice synthesis approach is introduced for speaker generation, which defines a speaker space directly within the model's parameter space via model editing. The method generates diverse unseen speakers, achieving comparable naturalness (MOS 3.81) and improved intelligibility (WER 2.20) compared to an x-vector baseline, and identified a controllable gender-dominant axis in the learned space.

Key insights

• A meaningful speaker space can be effectively defined and manipulated directly within the high-dimensional parameter space of a Deep Neural Network using model editing principles.
• The learned speaker space contains interpretable axes, such as a prominent gender-dominant axis, enabling direct control over specific speaker attributes.
• High-quality and diverse unseen speakers can be generated efficiently by modifying model parameters, without requiring external speaker embedding modules or complex multi-stage training.

Summary by alphaXiv.