Research Of Speech Enhancement Algorithm Based On Indoor Mobile Source

Multimedia technology and communication technology develop rapidly, people haveentered the era of intelligence, speech signal is the most direct and simple way, and plays animportant role. However, due to the impact of noisy environment and the process ofcommunication, the listener has not received the pure voice, but noisy speech, it influencespeople’s auditory effects and daily life seriously. Therefore, dealing with the noise of thevoice, improving communication quality is particularly important, this is the speechenhancement technology.Because the observed signal is complex, maybe it is stationary or moving, and thesurrounding noise is changing, room reverberation is also existence. Getting target signalbecomes difficult. Blind source separation and beamforming are two kinds of methods tosolve this problem, this thesis combines the two approaches. First beamforming is used aspre-processing of speech enhancement to weaken the influence of reverberation. Then, useindependent component analysis to separate the signals, after that, still has crosstalkcomponent, it need post-processing to improve the performance of speech enhancement.This thesis mainly for indoor mobile speech, the main work as follows:Firstly, for indoor environment, walls can reflect speech signal, multiple reflections aresuperposed to form reverberation, not only affects people’s hearing, but also cause crosstalklater. This thesis use beamforming as pre-processing, aimed at the target speech source byadaptive beamforming, which is the main ingredient in the output, thus have highersignal-to-noise ratio, and realize enhancing the useful signal and weaken interference.Secondly, the actual speech signal model is closer to the convolutive mixture model, thesource signal is constantly moving, therefore, the mixed system in the separation process istime variant. The thesis introduces a method combined independent component analysis (ICA)with binary time-frenquency masking. Introduce frequency domain ICA, when using onlinealgorithm, the computation and convergence speed is slow, a batch algorithm based on theblock data is introduced to compensate for these shortcomings. Speech separation simulationexperiment is performed at last.Thirdly, there is crosstalk component of the separated speech signal, post-processing isneeded. analyze the component of crosstalk and estimate the crosstalk model, using adaptivemethod to purify it, Compare the performance of post-processing of online algorithm andbatch algorithm, verify the feasibility of the algorithm, finally, the thesis evaluate the wholespeech enhancement system.