Using an Adaptive NLMS and Echo Suppressor for Real-Time Acoustic Processing on an Embedded System

Abstract

This paper presents a scheme of Normalized Least Mean Square (NLMS) adaptive filters, implemented in time and frequency domains, for real-time Acoustic Echo Cancellation (AEC) applied on the embedded system, in this case is Raspberry Pi. Recognizing the inherent limitations of linear adaptive filtering in complex acoustic paths, these implementations are augmented with a frequency-domain Residual Echo Suppressor (RES). The primary objective is to evaluate the efficacy of these hybrid architectures for voice agent deployment on edge computing platforms, specifically the Raspberry Pi 5. Our experimental framework evaluates performance across three distinct scenarios: an external hardware setup (asymmetric mic/speaker distances), an integrated hardware setup (laptopnative components), and a pre-recorded reference scenario from the public NKF-AEC repository. Performance is analyzed with respect to physical transducer proximity-a critical factor in environments where users interact via open speakers rather than head-worn peripherals.