DSENet: Directional Signal Extraction Network for Hearing Improvement on Edge Devices.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Microphone arrays capture complex soundscapes with reverberation and multiple sources.
• Traditional beamforming methods often suffer from crosstalk and latency issues.
• Existing spatially constrained methods may not be practical for real-time applications.

Purpose of the Study:

• To propose a novel directional signal extraction network (DSENet) for real-time audio processing.
• To address the limitations of existing methods in handling reverberant and multi-source environments.
• To enable practical hearing improvement solutions on edge devices.

Main Methods:

• DSENet utilizes a computationally efficient, low-distortion linear model in the time domain.
• The network is designed for low-latency, real-time performance.
• It extracts signals from a specified directional region of interest, handling multiple sources simultaneously.

Main Results:

• DSENet outperforms oracle beamformers and state-of-the-art low-latency causal speech separation methods.
• The system achieves a remarkably low latency of only 4 milliseconds.
• Successful real-time deployment on a smartphone demonstrates practical viability.

Conclusions:

• DSENet offers a practical and effective solution for directional signal extraction in reverberant conditions.
• Its low latency and computational efficiency make it ideal for hearing assistance on edge devices.
• The proposed method circumvents crosstalk issues inherent in traditional beamforming.