Speech Perception Improvement Algorithm Based on a Dual-Path Long Short-Term Memory Network

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Deep learning speech enhancement often focuses on time-frequency representations.
• Conventional methods can cause speech distortion due to domain-specific processing limitations.

Purpose of the Study:

• To develop a novel speech enhancement model addressing limitations of single-domain approaches.
• To improve speech quality and intelligibility by integrating temporal and spectral feature extraction.

Main Methods:

• A dual-path deep learning network was designed, incorporating both time and time-frequency domains.
• The time path models semantic features from the waveform, while the time-frequency path refines spectral details.
• Long Short-Term Memory (LSTM) networks were utilized to model mask functions for enhancing temporal and spectral features.

Main Results:

• The proposed dual-path LSTM network demonstrated superior performance compared to single-domain methods.
• Consistent improvements were observed in both objective and subjective evaluations of speech quality.
• Enhanced intelligibility was a key outcome of the dual-path approach.

Conclusions:

• The dual-path architecture effectively captures complementary speech information from different domains.
• This model offers a more comprehensive and robust solution for speech enhancement challenges.
• The findings suggest a promising direction for future deep learning-based audio processing research.