C-EMDNet: A Nonlinear Morphological Deep Framework for Robust Speech Enhancement.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Traditional speech enhancement methods often struggle with complex noise due to fixed time-frequency representations.
• The need for adaptive and structure-preserving noise reduction techniques is critical for improving speech intelligibility.

Purpose of the Study:

• To introduce C-EMDNet, a nonlinear speech denoising approach.
• To leverage Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and deep learning for enhanced speech enhancement.

Main Methods:

• Utilized CEEMDAN to decompose speech signals into intrinsic mode functions (IMFs), treating them as a morphological latent space.
• Employed a U-Net-like deep convolutional network to estimate mode-wise masks for noise suppression.
• Operated directly in the IMF domain, preserving speech's multi-scale structural information.

Main Results:

• C-EMDNet demonstrated superior performance compared to classical denoising algorithms.
• The proposed method outperformed competitive deep learning-based speech enhancement baselines.
• Preservation of harmonic and formant structures was achieved through selective noise suppression.

Conclusions:

• Nonlinear morphological representations offer a promising alternative framework for speech enhancement.
• C-EMDNet effectively addresses limitations of conventional methods by adapting to signal characteristics.
• The approach shows significant potential for real-world speech processing applications.