Speech preprocessing and enhancement based on joint time domain and time-frequency domain analysis.

Area of Science:

• Signal Processing
• Audio Engineering
• Machine Learning

Background:

• Traditional speech enhancement methods operating in the time-frequency domain often distort speech signals when separating them from noise.
• Differentiating between speech and noise remains a challenge for existing time-frequency domain techniques.

Purpose of the Study:

• To develop an improved speech enhancement approach by integrating time and time-frequency domain processing.
• To introduce the TTF-W-Net, an enhanced Wave-U-Net module for noise suppression.

Main Methods:

• A novel TTF-W-Net module was developed, improving upon the Wave-U-Net architecture.
• Experiments involved integrating Wave-U-Net and TTF-W-Net as preprocessing networks into baseline methods like Phase, FullSubNet+, and DB-AIAT.
• The TIMIT speech and NOISEX-92 noise datasets were utilized for performance evaluation.

Main Results:

• The TTF-W-Net preprocessing network demonstrated superior performance compared to the baseline Wave-U-Net.
• TTF-W-Net achieved a 15.7% improvement on the Perceptual Evaluation of Speech Quality (PESQ) metric.
• Integrating TTF-W-Net as a preprocessing step significantly enhanced the performance of baseline speech enhancement methods.

Conclusions:

• The TTF-W-Net preprocessing network offers an effective solution for advanced speech enhancement.
• Combining time and time-frequency domain processing via TTF-W-Net leads to more robust noise suppression.
• The proposed method shows significant potential for improving the clarity of noisy speech signals.