On Cross-Corpus Generalization of Deep Learning Based Speech Enhancement

Area of Science:

• Speech processing
• Machine learning
• Signal processing

Background:

• Deep neural networks (DNNs) are standard for speech enhancement, generalizing well to new noises and speakers when trained broadly.
• However, DNNs exhibit poor generalization to new speech corpora, particularly in low signal-to-noise ratio (SNR) environments.

Purpose of the Study:

• Investigate the causes of poor cross-corpus generalization in DNN-based speech enhancement.
• Identify effective techniques to improve generalization performance, especially under low SNR conditions.

Main Methods:

• Analyzed DNN generalization failures, identifying channel mismatch as a primary cause.
• Evaluated traditional channel normalization and publicly available datasets for generalization potential.
• Investigated the impact of frame shift size in short-time Fourier transform (STFT) processing.

Main Results:

• Channel mismatch significantly hinders DNN generalization to new speech corpora.
• Traditional channel normalization methods are insufficient for improving cross-corpus generalization.
• A smaller frame shift in STFT processing notably enhances cross-corpus generalization.
• One specific public dataset demonstrated superior potential for generalization.

Conclusions:

• Channel mismatch is a critical factor limiting DNN speech enhancement generalization.
• Optimizing training data and employing smaller STFT frame shifts are key strategies for improved cross-corpus generalization.
• The proposed methods collectively enhance speech intelligibility and quality on unseen data.