End-to-End Deep Convolutional Recurrent Models for Noise Robust Waveform Speech Enhancement.

Area of Science:

• Speech processing
• Deep learning
• Signal processing

Background:

• End-to-end deep learning (E2E-DL) models are popular for speech enhancement due to their simple design.
• While effective, creating resource-efficient and compact models for real-time processing remains a challenge.
• Modeling sequential and local characteristics of speech signals is crucial for improving E2E model performance.

Purpose of the Study:

• To present resource-efficient and compact neural models for end-to-end, noise-robust, waveform-based speech enhancement.
• To integrate Convolutional Encode-Decoder (CED) and Recurrent Neural Networks (RNNs) within the Convolutional Recurrent Network (CRN) framework for diverse speech enhancement systems.

Main Methods:

• Developed novel Convolutional Recurrent Network (CRN) models combining Convolutional Encode-Decoder (CED) and Recurrent Neural Networks (RNNs).
• Trained and tested models using various noise types, speakers, and datasets (LibriSpeech, DEMAND).
• Evaluated model performance based on quality, intelligibility, trainable parameters, model complexity, and inference time.

Main Results:

• Proposed models achieved improved speech quality (31.61%) and intelligibility (17.18%) compared to noisy speech.
• Demonstrated reduced model complexity and inference time compared to existing recurrent and convolutional models.
• Cross-corpus analysis confirmed the generalization capability of the proposed end-to-end speech enhancement (E2E SE) models.

Conclusions:

• The developed CRN-based E2E SE models offer a superior balance of performance and efficiency.
• These models provide significant improvements in speech quality and intelligibility with reduced computational demands.
• The findings highlight the effectiveness of integrating CED and RNNs for noise-robust speech enhancement.