Adaptive Filtering Techniques for Real-Time Audio Signal Enhancement in Noisy Environments

Abstract

Real-time audio processing systems in real world environments must contend with the fact that they will often encounter non-stationary noise sources that degrade intelligibility of the speech and overall audio quality. Noise conditions are common in mobile communication, hearing aids, teleconferencing, and human-machine interfaces, but since noise dynamics are rapidly varying, traditional fixed filtering techniques fail to adapt to the changing noise conditions. This paper carries out a systematic investigation of the performance of the Least Mean Squares (LMS), Normalized LMS (NLMS), Recursive Least Squares (RLS), and Affine Projection Algorithms (APA) for real time audio signal enhancement. The convergence behavior, noise suppression effectiveness, computational overhead and the embedded feasibility of these algorithms are analyzed. Through experiments with rigorous noisy speech data experiments, we show that while RLS has the least amount of noise and most signal fidelity, its computational complexity may constrain its use in resource constrained systems. On the other hand, NLMS comes out with a good compromise, allowing for reliable performance, with a relatively low latency and power requirements. In addition, we design a hybrid adaptive filtering scheme combining NLMS and spectral subtraction algorithms to cope with instantaneous acoustic noises in real time. The approach is validated on embedded hardware, the Raspberry Pi and Jetson Nano, with low inference time yet achieving tremendous gains in SNR and perceptual speech quality metrics. Adaptive filtering and its hybridisation with spectral analysis is confirmed to be a promising direction of achieving near real time audio enhancement in a noisy environment across a broad spectrum of industrial and consumer applications.