Determined Blind Signal De-Noising for Enhanced Communication Using Machine Intelligence

Abstract

A hands-free speech recognition system and a hands-free telecommunication system are essential for realizing an intuitive, unconstrained, and stress-free human-machine interface. In real acoustic environments, however, the speech recognition and speech recording are significantly degraded because one cannot detect the user’s speech signal with a high signal-to-noise ratio (SNR) owing to the interfering signals such as noise. In this thesis, blind source separation (BSS) algorithm and Artificial Neural Networks (ANN) are applied to overcome this problem. Artificial intelligence (AI) is the intelligence behaviour exhibited by machines or software. Intelligent Machine, therefore, is a system that perceives its environment and takes actions that maximize its chances of success. Radial Basis Function network is used in this research work.

Independent Component Analysis (ICA); a statistical signal processing technique having emerging and new practical application areas, such as blind signal separation and analysis of several types of data for feature extraction, is used as a preprocess to ANN. In blind separation, ICA algorithm separates the independent sources from their mixtures by measuring non-Gaussian variables of data. Blind ICA is a common method used to identify artefacts and interference from their mixtures and is applied in fields such as electroencephalogram (EEG), magnetoencephalography (MEG), and electrocardiogram (ECG). Therefore, based on these valuable applications, ICA is implemented for real-time signal processing like in hands free communications systems. The ICA-based BSS can be classified into two groups in terms of the processing domain, i.e., frequency-domain ICA (FDICA) and time-domain ICA (TDICA). This thesis implements time domain Independent Component Analysis (ICA) to separate signal mixtures. Blind ICA also acts as a preprocess for RBF network such that the network complexity is reduced.

This thesis, therefore, presents ICA-Radial Basis Function (ICA-RBF) based on maximum entropy which performs separation of mixed signals and generalization of input signals. The proposed algorithm for blind source system, also maximizes Signal-to-Interference Ratio (SIR) and Signal-to-Distortion Ratio (SDR) of the extracted signals. This research work specifically emphasizes on information-theoretical approach, filtering and the associated adaptive nonlinear learning algorithm, focusing on a numbers of signals: audio signals and digital communications signals (polar non-return to zero signals).The results have shown that a RBF network with ICA as an input pre-process has not only a better generalization ability to the one without pre-processing, but also the former’s performance converges much faster. To verify the proposed algorithm, MATLAB simulations are also performed for both off line signal processing and real-time signal processing show that the proposed method gives better Signal to interference Ratio and Signal to Distortion Ratio than early ICA techniques without involving neural networks. MATLAB implementation codes are included as appendices.