CROSS-DOMAIN DIFFUSION BASED SPEECH ENHANCEMENT FOR VERY NOISY SPEECH.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Machine Learning

Background:

• Deep learning has advanced speech enhancement, yet struggles with low signal-to-noise ratio (SNR) and nonstationary noise.
• Robust speech enhancement is crucial for applications in adverse acoustic environments.

Purpose of the Study:

• To develop a more robust deep learning model for speech enhancement, particularly effective in extremely noisy conditions.
• To improve speech recovery in low SNR scenarios by incorporating diffusion-based generative learning.

Main Methods:

• A frequency-domain diffusion-based generative module was developed.
• This module utilized an enhanced signal from a time-domain supervised enhancement module as auxiliary input.
• The model was trained to recover clean speech spectrograms.

Main Results:

• The proposed model demonstrated superior speech enhancement performance compared to strong baselines.
• Significant improvements were observed in extremely noisy conditions with SNR levels of -5 dB and -10 dB.
• Experiments were conducted on the TIMIT dataset.

Conclusions:

• Integrating diffusion-based learning enhances the robustness of deep learning speech enhancement models.
• The proposed approach effectively addresses challenges posed by low SNR and nonstationary noise.
• This method offers a promising direction for future research in robust speech processing.