Advanced Two-Sensor Adaptive Algorithms for Noise Reduction in Dispersive and Sparse Acoustic Environments

Abstract

In this thesis, we address the problem of acoustic noise reduction and speech enhancement in modern telecommunication systems using two-sensors adaptive filtering and intelligent learning-based approaches. The research focuses on the development of advanced two-sensor adaptive algorithms capable of efficiently separating speech from noise in dispersive and sparse acoustic environments, where conventional methods often fail. In the first part of this project, we propose a novel Neural Network-based Variable Step-Size Feed-forward NLMS (NN-V-FNLMS) algorithm. This approach integrates a simple neural network to dynamically estimate the adaptation step-size, thus overcoming the inherent trade-off between fast convergence and low steady-state error found in conventional algorithms. A Voice Activity Detector (VAD) is also incorporated to control filter updates and improve computational efficiency. To further enhance robustness and adaptability, we propose a contribution that introduces an advanced Deep Learning Variable Step-Size Feed-forward NLMS (DL-VSS FNLMS) algorithm. This model employs a Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) layers to predict optimal step-size parameters based on rich acoustic features, including Mel-Frequency Cepstral Coefficients (MFCCs), GammatoneCepstral Coefficients (GTCCs), and spectral representations on the ERB, Bark, and Mel scales. This deep model dynamically adjusts the adaptation process in complex acoustic conditions, achieving superior performance in both dispersive and sparse environments. Experimental evaluations demonstrate that the proposed algorithms significantly improve convergence speed, stability, and speech quality, while ensuring better noise suppression compared to conventional LMS/NLMS and classical VSS algorithms. The results confirm that the integration of neural networks and deep learning into two-sensor adaptive filtering offers a powerful and flexible solution for real-world acoustic noise reduction and speech enhancement.