Real-time single-channel deep neural network-based speech enhancement on edge devices

Area of Science:

• Artificial Intelligence
• Signal Processing
• Acoustics

Background:

• Speech enhancement (SE) is crucial for clear audio communication.
• Existing methods often struggle with real-time processing on edge devices.
• Deep learning offers potential for advanced SE algorithms.

Purpose of the Study:

• To develop a real-time, single-channel speech enhancement model using deep neural networks.
• To implement and evaluate the model on a smartphone for practical usability.
• To compare the proposed method against conventional and deep learning-based SE techniques.

Main Methods:

• A hybrid deep neural network architecture combining convolutional neural network (CNN) and recurrent neural network (RNN) layers was designed.
• The model processes the noisy speech magnitude spectrum on a frame-by-frame basis.
• Implementation on a smartphone (edge device) for real-time performance evaluation.

Main Results:

• The proposed deep learning model demonstrated effective real-time speech enhancement on a smartphone.
• Objective metrics, including Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI), were used for comparison.
• Subjective listening tests confirmed superior performance compared to baseline SE methods.

Conclusions:

• The developed CNN-RNN architecture provides a viable solution for real-time speech enhancement on edge devices.
• The model achieves significant improvements in speech quality and intelligibility.
• This research paves the way for enhanced mobile audio experiences.