Research Of Speech Signal Sparse Representation Method

Speech is the most common way of communication by far. In recent years, speech communications always combine with other multimedia communication mode, such as TV conference call, We-Chat etc. With the popularity of these applications, people not only demand for quantity of speech information increasingly, but also demand for the high quality of speech information. Simply get digital speech signal by the traditional sampling and quantify will occupy too much channel resources and require too much storage space when saving signals. Therefore, in the case of ensure the reliability of the speech communication as far as possible, how to compress the speech signal efficiently and reduce digital rate and occupied bandwidth become very important.Speech signal sparse representation can be an effective tool to decrease the digital rate and reduce the occupied bandwidth. In this dissertation, the speech signal sparse representation is studied. At first, the dissertation summarizes the sparse representation theory, while deep analysis two key technologies in sparse representation that sparse decomposition algorithm and the structuring of the sparse basis.1. Firstly, studied the K-SVD dictionary training algorithm and compared it with K-Means algorithm and MOD algorithm. K-SVD algorithm can training dictionary for a specific signal and get the sparse representation of signals at the same time. K-Means can be seen as a special K-SVD algorithm whose code size is one. The K-SVD algorithm use rank-1 approximation of error matrix kE instead of matrix inversion. This dissertation elaborates the K-SVD algorithm through analyzing the difference between these three algorithms.2. Based on linear representation method, this dissertation study the dictionary initialized problem of K-SVD algorithm. It’s necessary to know the size of dictionary at the beginning of the process based on K-SVD. Speech signal will suffer from over representation or under representation which impact the quality of signal representation significantly if you choose the unfit dictionary size. In view of the initial selection of dictionary size problems, this dissertation put forward a new dictionary training method based on the new BDS model. The method according to the relationship between dictionary size and sparse ratio established model for dictionary size. It was adaptive to speech signal choosing the proper initial dictionary and overcame the drawback of K-SVD method that set up dictionary size based on the experience. This dissertation applied this training dictionary method that added the BDS model to speech library of audio and video lab of Taiyuan University of Technology. The simulating experiments and the result are analyzed. Simulation results show that, based on BDS model speech signal dictionary construction method realized choose best dictionary size adaptively, and can improve the efficiency and stability of the dictionary training.3. According to the nonlinear feature of speech signal, this dissertation proposed a speech sparse representation method based on kernel dictionary. This article map speech signal to high-dimensional feature space, then linear represent the feature space signal, which imply nonlinear sparse representation of the speech signal. Algorithm introduced nuclear theory into calculating process to avoid mass calculation of higher dimensional space. The result shows that our method perform better than K-SVD when dealing with sparse represent non-linear signals. Also, this method can training dictionary for speech signals more effective than kernel MOD algorithm.