PVGAN: A Pathological Voice Generation Model Incorporating a Progressive Nesting Strategy.

Area of Science:

• Medical technology
• Artificial intelligence
• Speech processing

Background:

• Limited pathological voice data hinders accurate voice disorder diagnosis.
• Existing models lack detailed feature learning for pitch, timbre, and frequency components in pathological voices.

Purpose of the Study:

• To propose a PVGAN network for generating pathological voice data.
• To improve voice disorder diagnosis by increasing dataset size and learning detailed audio features.

Main Methods:

• Developed a PVGAN network with multiscale perceptual residual blocks and periodic discriminators.
• Implemented a progressive nesting strategy for generator-discriminator integration.
• Designed a latent mapping network for conditional voice feature generation.
• Optimized the loss function for enhanced model performance.

Main Results:

• Generated pathological voice data closely matched original data characteristics on the Saarbruecken Voice Database (SVD).
• The PVGAN network effectively learned multi-scale features and periodic patterns in audio signals.
• Dataset expansion using generated data led to improved accuracy in classification experiments.

Conclusions:

• The PVGAN network offers a viable solution for pathological voice data generation.
• This approach can significantly enhance the accuracy and applicability of voice disorder diagnosis systems.
• The method demonstrates potential for broad applications in medical diagnosis and related fields.