A Deep Ensemble Learning Method for Monaural Speech Separation.

Area of Science:

• Speech Processing
• Artificial Intelligence
• Machine Learning

Background:

• Monaural speech separation is crucial for clear audio in single-microphone environments.
• Deep neural networks (DNNs) have advanced speech separation but have limitations in contextual information and understanding optimization objectives.
• Existing DNN methods often use a fixed window length, limiting their ability to capture diverse speech characteristics.

Purpose of the Study:

• To propose a novel deep ensemble method, multicontext networks, for enhanced monaural speech separation.
• To address the limitations of single DNNs in leveraging contextual information and understanding optimization objectives.
• To systematically compare the effectiveness of predicting clean speech versus ideal time-frequency masks.

Main Methods:

• Developed two multicontext network architectures: one averaging outputs of DNNs with different window lengths, and another stacking DNNs with varied contexts.
• Each DNN in the stacked network processes concatenated acoustic features and expanded soft outputs from lower modules.
• Conducted extensive experiments on three speech corpora to validate the proposed method.

Main Results:

• The proposed multicontext networks significantly improve monaural speech separation performance.
• Demonstrated the effectiveness of ensemble methods in overcoming single DNN limitations.
• Found that predicting ideal time-frequency masks is more efficient for training data, while predicting clean speech offers better robustness to Signal-to-Noise Ratio (SNR) variations.

Conclusions:

• Multicontext networks offer a superior approach to monaural speech separation compared to single DNN models.
• The study provides valuable insights into the comparative advantages of different DNN optimization objectives in speech separation.
• The findings contribute to the development of more robust and effective speech processing technologies.