Speech Enhancement For The Telecommunication And Speech Recognition Systems

This thesis considers the problem of speech enhancement and noise reduction by adaptive filtering algorithms in the telecommunication and speech recognition systems.The presence of these systems in real noisy environments reduces their effectiveness and makes degradation in their performance.For example,for the telecommunication system,the background noise signals corrupt the transmitted speech signal and make degradation in its intelligibility and quality.For speech recognition systems,they make mismatch between the testing and training speech signals.Several single microphone enhancement techniques have been proposed in the past and are discussed here.Most of these techniques are based on the assumption of the noises stationarity but in the real life environment,this assumption is not widely verified which make these techniques affect the speech signal by many distortions.In addition,these techniques are limited by the underperformance at low input SNRs(<=5dB),noise levels and noise types.These limitations lead us to use the second microphone which allows getting signals characteristics and make us free from the assumption of the noise stationarity.However,the presence of two microphones provides two observation signals.The problem here is how can we separate the two source signals(speech and noise)by using only the observation signals without any aprior informations on the sources signals?.This problem is known under the name Blind Sources Separation(BSS)which is the most important problems in speech enhancement field.Therefore,several Two-Microphone Blind Sources Separation(TM-BSS)algorithms have been proposed to solve this problem.In this thesis,Firstly,a brief history and literature review about speech enhancement techniques are presented.Secondly,an exhaustive study about the theoretical basis of adaptive filtering algorithms are done in order to accomplish a deep understanding of the practical problems in this area.Then,three conventional TM-BSS algorithms are presented and analyzed before proposing our main contribution the novel algorithm.The proposed algorithm is called TwoMicrophones Reduce size Simplified Fast Transversal Filter(TM-RSMFTF)algorithm which is an innovative alternative of the classical algorithms previously proposed.This proposed algorithm is an outcome of the good combination between the well-known forward blind source separation structure and the adaptive filtering algorithm reduce size simplified fast transversal filter(RSMFTF).Finally,numerical simulations have been carried out under different conditions and situations,and the obtained results have shown the good performances and the effectiveness of this new TMRSMFTF algorithm in terms of the computational complexity and various objectives criteria such as Segmental SNR,System Mismatch and Segmental MSE.A fair comparative study with various TM-BSS algorithms has been presented as well in this thesis.Moreover,the conclusion has been drawn,and future works are proposed.