Dense CNN with Self-Attention for Time-Domain Speech Enhancement

Area of Science:

• Signal Processing
• Machine Learning
• Audio Engineering

Background:

• Time-domain speech enhancement is gaining traction for its ability to improve both speech magnitude and phase.
• Existing methods face challenges, particularly with magnitude-based loss functions.

Purpose of the Study:

• To propose a novel dense convolutional network (DCN) with self-attention for time-domain speech enhancement.
• To introduce a new loss function that addresses limitations of spectral magnitude-based losses.

Main Methods:

• Developed a DCN architecture with encoder-decoder structure, skip connections, dense blocks, and self-attention modules.
• Proposed a novel loss function based on enhanced speech magnitudes and predicted noise, ensuring phase enhancement.

Main Results:

• The DCN with the proposed loss function demonstrated superior performance in speech enhancement.
• The method significantly outperformed state-of-the-art approaches for both causal and non-causal speech enhancement tasks.

Conclusions:

• The proposed DCN and novel loss function offer a powerful solution for time-domain speech enhancement.
• This approach effectively enhances both speech magnitude and phase, leading to substantial improvements in speech quality.