A Lightweight Causal Sound Separation Model for Real-Time Hearing Aid Applications

Abstract

Real-time audio processing is crucial for hearing aid IoT applications, where low latency and efficiency are paramount. State-of-the-art models like Demucs achieve high signal-to-distortion ratio (SDR) but are unsuitable for real-time use due to their noncausal nature and high latency. This letter introduces a lightweight causal model tailored for real-time hearing aid applications, designed to minimize latency while maintaining acceptable SDR. The model was trained and evaluated on the MUSDB-18 dataset using established protocols. Performance metrics, including SDR and latency, were used to compare it against Demucs. Results show that while Demucs achieves higher SDR, the proposed model significantly reduces latency (9.42 ms compared to 52.25 ms), making it suitable for real-time IoT systems. This research demonstrates the potential of causal architectures in addressing the challenges of real-time audio processing for hearing aids and sets the stage for future improvements in SDR without compromising latency.