Statistical Model Of Speech Enhancement Algorithms

In real environment, the speech is often affected by outside noise which results in decline in the quality of speech communication and does not make the voice processing device perform normally to work. In order to reduce or eliminate mixed noise in the speech, the original signal will be restored in noisy speech to enhance the quality of voice communication. Speech enhancement algorithms are usually divided into three categories:one is in time domain; another is in transform domain and the other method. Transform domain algorithms can provide a maximum of energy compaction and help to eliminate the correlation between the speech components and noise components, which means that the noise of the noisy speech is removed easily. So transform domain algorithms become the focus of the research in the speech enhancement technology.Firstly, after many algorithms of speech enhancement are described, the paper considers a priori SNR estimation in speech enhancement. a momentum is added to reduce inaccurate estimation of decision-directed (DD), a new priori SNR estimation algorithm which improves the ability of tracking the instantaneous SNR and reducing the delay of DD algorithm is proposed. the two-step noise reduction technique solves the delay of DD approach, but it weakens the ability to eliminate background noise and musical noise,and the delay of DD method is still insufficient. In order to reduce the impact of the above issues on the performance, an improved two-step noise reduction algorithm is proposed in the paper, which compensates the error between the actual voice and its estimation to the output signal.Secondly, it addresses the algorithms of speech enhancement by considering speech model of Laplacian distribution in transform domain and gives the MMSE and ML estimators employing the Laplacian-Gaussian mixture model which produce better ability in noise reduction. We analyze the important of Laplacian factor for speech enhancement, and based on this research. a new estimated method using improved Laplacian model factor employing the Laplacian model is derived, which obtains an accurate estimates of factor and improve effectively the performance of the corresponding algorithm. And then, we discuss the shortcoming of the minimum statistics tracking algorithm and give two algorithms employing Gaussian model and Laplacian model, respectively. Simulation results show the better processing capabilities of the improved algorithms compared to that of the previous algorithm.Then, we make a study of the combination of speech model for speech enhancement. We discuss and analyze the distribution of the noise component and the voice component in the transform domain, and demonstrate that the statistical distribution of the actual voice and can be not expressed accurately by a single statistical model, it has different statistical distribution in different times and in different frequencies. Based on the above analysis, the combination of speech model is applied to speech enhancement, which contributes to estimate accurately the model of the actual voice. It ultimately provides a more ideal background for researching algorithms of voice-based statistical. Simulation results show that the proposed algorithm has better performance in different kinds of noises, providing a good theoretical basis for the further study of speech model.Finally, conclusions and future research works are described.