Monaural Speech Dereverberation Using Temporal Convolutional Networks with Self Attention.

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Human speech is often degraded by room reverberation in everyday listening environments.
• Reverberation significantly impairs the performance of speech processing systems compared to anechoic conditions.

Purpose of the Study:

• To propose a novel monaural (single-channel) speech dereverberation algorithm.
• To address the challenges posed by reverberation in speech processing.

Main Methods:

• Utilizing temporal convolutional networks (TCNs) with a self-attention mechanism.
• Incorporating a self-attention module for dynamic feature representation.
• Employing a TCN for non-linear mapping to anechoic speech magnitude spectrum.
• Using a 1-D convolution module for inter-frame magnitude smoothing.

Main Results:

• The proposed algorithm significantly improves objective speech quality metrics across diverse reverberant conditions.
• Demonstrated effectiveness in a wide range of reverberation times and room sizes.
• Showcased generalization to measured room impulse responses and real-world noisy-reverberant speech.

Conclusions:

• The developed algorithm offers a robust solution for monaural speech dereverberation.
• The approach shows excellent generalization capabilities across various acoustic scenarios and speakers.
• This method enhances speech intelligibility and quality in challenging acoustic environments.