Lightweight speech enhancement via learnable prior and Schrödinger bridge generative adversarial networka)

Area of Science:

• Artificial Intelligence
• Signal Processing
• Machine Learning

Background:

• Conventional Schrödinger bridge methods for speech enhancement face challenges with inefficient transport paths, high computational costs, and reduced quality in limited resource scenarios.
• Path crossings in existing models lead to computational inefficiencies and degraded performance.

Purpose of the Study:

• To develop an advanced speech enhancement framework that overcomes the limitations of conventional Schrödinger bridge methods.
• To improve speech quality, naturalness, and computational efficiency, particularly in resource-constrained environments.

Main Methods:

• Integration of a learnable prior module with a Schrödinger bridge generative adversarial network.
• Utilizing adversarial training and multi-scale loss functions for enhanced speech quality.
• Developing a nano-sized model (0.04M parameters) for efficient performance.

Main Results:

• The proposed framework significantly outperforms state-of-the-art methods in speech enhancement.
• Achieved superior performance across metrics like overall quality (OVRL), speech quality (SIG), background noise quality (BAK), and P808.MOS.
• Demonstrated robust and competitive results even with a compact, nano-sized model in resource-limited conditions.

Conclusions:

• The synergistic combination of learnable prior and adversarial modeling effectively enhances speech quality and efficiency.
• The learnable prior module is crucial for minimizing path crossings and optimizing transport paths.
• This approach offers a computationally efficient and high-performing solution for speech enhancement, especially in edge computing scenarios.