Speech Enhancement using Time Domain Adaptive Wiener Filtering with LMS and NLMS Algorithm

Abstract

Speech signals often suffer from degradation due to additive noise, which significantly affects their clarity and intelligibility. To address this issue, an adaptive speech enhancement technique based on the Wiener Filter optimized with the Normalized Least Mean Squares (NLMS) algorithm is proposed. This method operates in the time domain, allowing real-time adaptation to the non-stationary characteristics of speech. Various types of real-world noise—such as traffic, wind, and household appliance noise—are artificially introduced into clean speech recordings for evaluation. The system is tested using different step-size (μ) values of 0.1, 0.2, and 0.3 in the LMS adaptation to analyze convergence behavior and noise suppression performance. Unlike traditional frequency-domain approaches, the time-domain implementation ensures fast convergence and efficient noise suppression while preserving key speech features. The performance of the system is quantitatively assessed using the Mean Squared Error (MSE) metric. Experimental results show that the NLMS-based Adaptive Wiener Filter with LMS adaptation at μ = 0.1, 0.2, and 0.3 provides significant noise reduction and enhanced speech quality. Due to its robustness and low computational complexity, the proposed method is highly suitable for practical applications in telecommunications, hearing aids, and voice-controlled systems.