Learning Complex Spectral Mapping with Gated Convolutional Recurrent Networks for Monaural Speech Enhancement.

Area of Science:

• Speech processing
• Machine learning
• Signal processing

Background:

• Speech phase is crucial for perceptual quality but difficult to estimate directly.
• Existing methods like magnitude spectral mapping and complex ratio masking have limitations.

Purpose of the Study:

• To propose a novel gated convolutional recurrent network (GCRN) for complex spectral mapping in monaural speech enhancement.
• To improve both speech intelligibility and perceptual quality by simultaneously enhancing magnitude and phase.

Main Methods:

• Developed a causal gated convolutional recurrent network (GCRN) inspired by multi-task learning.
• Applied complex spectral mapping to estimate real and imaginary spectrograms of clean speech from noisy speech.

Main Results:

• The proposed GCRN significantly outperformed a convolutional neural network (CNN) for complex spectral mapping.
• Achieved higher STOI (Speech Transmission Quality Index) and PESQ (Perceptual Evaluation of Speech Quality) scores compared to other methods.
• Demonstrated effective phase estimation capabilities.

Conclusions:

• Complex spectral mapping with the proposed GCRN is a highly effective approach for monaural speech enhancement.
• The GCRN architecture offers substantial improvements in speech intelligibility and quality.
• This method provides a viable solution for accurate speech phase estimation.