Logo image
Back
Real-Time Environment-Specific Deep Denoising for Edge-Assisted Hearing Aids
Conference proceeding

Real-Time Environment-Specific Deep Denoising for Edge-Assisted Hearing Aids

Yumna Anwar, Ian Pope, Evan Finken, Yu-Hsiang Wu, Steve Goddard and Octav Chipara
Proceedings (IEEE International Conference on Healthcare Informatics. Online), pp.1017-1026
06/01/2026
DOI: 10.1109/ICHI69079.2026.00127

View Online

Abstract

We propose an end-to-end framework for real-time deep speech enhancement on embedded hearing aids that combines environment-specific model training, edge-assisted deployment, and a custom C-based inference engine. Instead of a single large universal model, a paired edge device dynamically deploys lightweight noise-specific models (e.g., Car, Cafe) to the hearing aid as the acoustic environment changes. Across 13 environments, compact specialized models (H = 8) match the objective quality (PESQ) of substantially larger universal models (H = 24), while meeting strict FDA latency constraints. A custom C inference engine achieves an end-to-end latency of 11.28 ms on the Portable Hearing Laboratory (PHL) platform. A user study with 30 listeners with mild-to-moderate hearing loss demonstrates a significant preference for environment-specific models over both universal DNNs and traditional signal-processing baselines.
 Noise Noise Reduction and real-time processing Auditory system deep learning Denoising Hearing aids Modeling neural networks Printing Real-time systems Speech enhancement Timing Training

Details

Metrics

1 Record Views
Logo image