Spectral network based on lattice convolution and adversarial training for noise-robust speech super-resolution

Area of Science:

• Signal Processing
• Artificial Intelligence

Background:

• Speech super-resolution enhances low-resolution audio to high-resolution.
• Existing models struggle with real-world noise and flexible sampling rates.
• Robustness in practical applications is a key challenge.

Purpose of the Study:

• To develop a noise-robust speech super-resolution model adaptable to flexible input sampling rates.
• Introduce Super Denoise Net (SDNet) for practical, real-world audio enhancement.

Main Methods:

• Designed Super Denoise Net (SDNet) incorporating gated and lattice convolution blocks.
• Utilized frequency transform blocks for long frequency dependency modeling.
• Employed a multi-scale discriminator for multi-adversarial loss training.

Main Results:

• SDNet demonstrates superior performance compared to state-of-the-art models.
• Achieved significant improvements in noise-robust speech super-resolution.
• Validated effectiveness across multiple test sets.

Conclusions:

• SDNet offers a robust solution for real-world speech super-resolution challenges.
• The model's design effectively handles noise and variable sampling rates.
• Indicates a significant advancement in practical audio enhancement technology.