Speech signal enhancement based on deep learning in distributed acoustic sensing

Area of Science:

• Acoustics
• Signal Processing
• Artificial Intelligence

Background:

• Speech signal fidelity in Distributed Acoustic Sensing (DAS) systems is severely degraded by random noise, impacting measurement accuracy.
• Existing methods struggle to effectively enhance speech signals within noisy DAS environments.

Purpose of the Study:

• To develop and evaluate a deep learning technique for enhancing speech signal recognition and reconstruction in DAS systems.
• To quantitatively assess the noise reduction and signal quality improvement achieved by the proposed method.

Main Methods:

• A novel Complex Convolution Recurrent Network (CCRN) algorithm utilizing complex spectral mapping was designed for speech signal enhancement.
• The performance of the CCRN was theoretically and experimentally validated against speech signals with and without the enhancement method in a DAS system.

Main Results:

• The proposed CCRN method significantly suppressed random noise in DAS-collected speech signals, attenuating noise intensity by approximately 20 dB.
• The Scale-Invariant Signal-to-Distortion Ratio (SI-SDR) was improved by an average of 51.97 dB, demonstrating superior performance compared to other enhancement techniques.
• Enhanced information identification and recognition capability of speech signals were observed.

Conclusions:

• The developed deep learning approach effectively achieves high-fidelity and high-quality speech signal enhancement in DAS systems.
• This method represents a significant advancement for high-performance DAS systems in practical applications.
• The CCRN algorithm shows promise for robust speech processing in challenging acoustic environments.