Enhancing speech separation performance utilizing various wavelet coefficients

Area of Science:

• Signal Processing
• Machine Learning

Background:

• Speech separation performance degrades in real-world noisy conditions due to feature distortion.
• Traditional speech separation features lack robustness in practical acoustic environments.
• Wavelet transform (WT) shows potential beyond classification for speech separation.

Purpose of the Study:

• To explore the efficacy of wavelet coefficients in improving speech separation models.
• To evaluate the impact of discrete wavelet and wavelet packets on speech separation.
• To integrate wavelet scattering (WS) coefficients into speech separation despite lacking an exact inverse transform.

Main Methods:

• Integrating discrete wavelet and wavelet packets into model training.
• Incorporating wavelet scattering coefficients into the loss function to address inverse transform limitations.
• Evaluating model performance using metrics like scale invariant-signal to distortion ratio, mean opinion score, and short time objective intelligibility.

Main Results:

• Wavelet-based models demonstrate superior performance and resilience in noisy conditions.
• Integrating WS coefficients significantly enhances speech separation accuracy.
• Wavelet-based methods outperform other approaches in key objective and subjective speech quality metrics.

Conclusions:

• Wavelet coefficients are effective for enhancing speech separation in challenging acoustic environments.
• Wavelet scattering integration provides a novel approach to improve separation accuracy.
• Wavelet coefficients represent state-of-the-art solutions for robust speech separation.