Research On Speech Feature Vectors Vector Quantization Algorithm Based On The Genetic Algorithm

Vector quantization is an efficient method for data compression. Its merit is high compression rate and easy decoding method. Vector quantization is used widely, such as compression coding and transfer in real time, store and transfer of radar image and military map, compression of digital TV signal and DVD, compression of medical image, compression and storage of data through internet, speech coding, speech recognition and so on. The research of vector quantization is important for the science and economical development.The problem of vector quantization for the speech feature vectors is investigated in this paper. The main research aim is to find the optimization code book. As the classic LBG method can't assure to find the global optimization code book, genetic algorithm is utilized for the optimization of code books. This method is helpful to improve the quantization quality and reduce the distortion measure.The basic principle of vector quantization is described as below. The input vector is replaced by the proper code word for the transfer and storage. Only simple search is needed at the coding stage. There are three key techniques in the vector quantization. They are code book design, search of code word and assignment of code index. The most important part is design of the code book. It will determine the performance of quantization. LGB method that depends on grads-descending method can not assure global optimization of the code book. Therefore genetic algorithm is used for the global search of the optimization code book. Training sample set of the quantization is speech feature vectors. They are Mel frequency cepstrum coefficients and linear prediction cepstrum coefficients. Process of vector' quantization of the speech feature vector using genetic algorithm is as below. The speech signal is predisposed first including sampling, pre-emphasizing, short-time framing, and speech endpoints detection. Then feature vectors of the speech section are obtained. The global optimization code book is got through genetic operation for the initial code books set.The speech signal is sampled first. The analog speech signal is turned into digital speech signal in time and amplitude. The average power spectrum is influenced by glottis inspiring and radiation between the nose and mouth. So components of high frequency are attenuated. Pre-emphasizing is needed to get a flat spectrum. The pre-emphasizing zeros and radiation zeros will get rid of influence of glottis wave. The speech signal will contain the information of track only. It is helpful for analyze of the spectrum and tack. Speech and noise have different energy. Usually energy of speech segment is bigger than that of noise segment. When signal-noise-ratio is high enough, the speech will be divided into voiced-segment and unvoiced-segment using short-time energy or short-time amplitude.The speech signal includes much redundant information. The speech is often replaced by feature to reduce the amount of data and quicken up processing speed. The wave in time-domain varies with time quickly. When different people make the same speech of the same speech, wave of time-domain is different.Speeches sample of the same text will exhibit much same characteristics after frequency domain analyze. Most of speech features come from frequency domain. The two features used in this paper are Mel frequency cepstrum coefficients and linear prediction cepstrum coefficients. linear prediction cepstrum coefficients is a orthogonal transformation of the linear prediction coefficients. It reflects the characteristics of the track. Mel frequency cepstrum coefficients reflects the nonlinear characteristics of the human hearings to frequency of the sound.Training vector sample set is trained to get the code book after speech feature vectors are obtained. Distortion measure needs to be fixed first. And then training vector sample set is divided into several classes. The multi-dimensional space is divided into several subspaces whose center is code word. Results of classic LBG method are only local minimums. The whole distortion is a function of all the code word. As the distortion function is not convex, the function has local minimums and global minimums. Property of the iterative algorithm depends on the initial value of the code words. As a result, code book developed by LBG method can not assure that the objective function reaches the global minimum.In order to overcome the shortcomings of the LBG vector quantization, genetic algorithm of global searching is proposed in this paper to get the optimization code book. Research of genetic algorithm began at sixties in twentieth century. It is widely used in the field of optimization computation. Compared with other optimization methods, genetic algorithm has many merits, such as it only needs adaptation function but not derivatives of the objective function. Genetic algorithm can be used for kinds of optimization problems. It could search many regions to overcome the shortcoming of local minimum in grads-descending method. Genetic algorithm could get the global optimization code book. If the objectives of genetic algorithm are chosen as binary string, it is easy for the hardware implementations. A code book is chosen as the individual. Not all the individuals are determined using the random choosing method. One individual is obtained through splitting method. As there is a sound code book in the initial individuals, the convergence speed will be accelerated. Choosing operation is implemented by rotary table method and sorting method. A half is chosen by rotary table method, and the other half is chosen by sorting method. Crossing operation is an important operation in the genetic algorithm. It reassembles individual genes to produce more excellent offspring.Crossing operation is two point crossing, crossing position is chose randomly. In the experiment, it is found that offspring individuals are not better than the father generation. It makes the optimization algorithm converge slowly even not converge. So the efficiency grading is introduced to make the colony evolves to the better direction. Efficiency grading could get the offspring excellent than father generation. And the excellence degree could be under control. Aberrance operation is also used to produce more excellent gene and individual and make the colony abundant in individuals. Efficiency grading is also performed after aberrance operation.Vector quantization experiment of speech feature vectors indicates that average quantization distortion of genetic algorithm based method is smaller than average quantization distortion of LBG method. In vector quantization experiment of the 12-dimension linear prediction cepstrum coefficients, average quantization distortion of genetic algorithm based method is 0.988, while average quantization distortion of LBG method is 2.052. Average quantization distortion of genetic algorithm based method is smaller than average quantization distortion of LBG method when size of code book is 32, 64, 128.