Hidden Markov Model (HMM) method has now become the mainstream technology of speech recognition, which is quickly identified in speech recognition, and has a high recognition rate. In HMM, it is divided into discrete HMM (DHMM) and continuous HMM (CHMM). Since the CHMM is directly in the frame speech feature vector itself is a sequence of observation, rather than the vector is quantified as a viewing symbol, the CHMM is superior to the identification accuracy of DHMM. However, since the CHMM parameter is more, the traditional training method uses iterative method, first assumes that the initial value is trained with the observation sequence of the voice signal, that is, the valuation is purified according to certain methods, and is purified. The valuation should be further purified until there is no further improved room, reaching a partial optimum value. Traditional training methods do not guarantee training to get the best solution in the whole domain, and the time required for training is very huge. This paper focuses on the gene algorithm [4, 5], and constructed chromosomes according to the characteristics of CHMM, and CHMM is trained with gene algorithm. The characteristics of gene algorithm itself make training results tend to the most domain optimal solution. At the same time, since only the VITERBI algorithm is required to calculate the relevant probability of a certain CHMM model with the Viterbi algorithm, the algorithm can be used as a gene algorithm, so the algorithm can improve the training speed of CHMM.
1 Theoretical Foundation of Algorithm
The gene is a biological concept, which introduces the gene algorithm into HMM training, because the training process of HMM is actually an iterative purification of the HMM model once in a particular range, selecting the optimal model. This is similar to each other, and the phenomenon of the survival of the fittest is similar. The genetic gene of the biological gene is contained in the chromosome. The chromosome is always paired, and the chromosome of the parent organisms each copy its own genetic pass to the child. After a certain cross, gene recombination forms the chromosome of the next generation of organisms, from the parent The genetic characteristics can be reflected and retained in the progeny. At the same time, there is a certain gene mutation while genetic. Gene mutations cause organism mutations, breaking the old balance, breaking through the old gene activity area, has a big impact on the evolution of species. The driving force of biological evolution comes from genetics and choices, whether normal gene recombination, or genetic mutation of burst-free, can control harmful elimination of sub-genes, but will only be beneficial to retain, so that the organism Evolution in a good direction. For better genes, the generation of generations constantly makes the genes of the child converge in a single genetic form, which is the optimal solution in a particular optimization problem. From the perspective of mathematics, it can be simply believed that the genetic recombination makes the genetic gene tends to be partially optimal, and the gene mutation makes the sub-gene break through the part of the part, and has passed many generations of genetics and options to achieve the domain optimal solution. The essence of the traditional CHMM training algorithm is to select a CHMM model as the initial value, which is to select the initial state vectors π. The output probability density function Bj (o) = σcjk n (o, μJk, ujk) of each state is calculated together with the observation sequence, and obtain a new, better than the old CHMM estimate. Model, repeated iteration, until local optimal solution. Several different initial values can be used, and it is desirable to reach a better optimum value. The introduction of the gene algorithm into CHMM is based on a problem based on the problem of being a constrained point of choice of CHMM as a specific domain. CHMM status transfer matrix A and output probability density function The sum of each line of each line of the C matrix of the C matrix is 1.0, which can be considered as a constraint condition for optimizing problems. If an initial value is selected, it is not selected to select a set of initial values distributed in different regions, which tends to be the best solution in the whole domain, then ultimately the same Training to CHMM can be completed.
2 gene algorithm
In nature, the driving force of biological evolution comes from genetic and selection. In the gene algorithm, the main operation is to simulate genetic genetically recombinant and gene mutations, as well as simulation of natural selection sample selection. Define suitable functions F (AI) based on the mathematical model of the problem to be optimized. Among them, AI is a chromosome, and the function F (AI) is the distance between the chromosome and the target function, or to determine the basis of the geographical strength. For each gene, the suitable function of all chromosomes is calculated, and the sort selection of a certain number of superior chromosomes as a parent sample for the production of the next generation. In the natural boundaries, the chromosomes were separated, and the chromosome was separated and restructured. Figure 1 is a schematic diagram of a multi-point cross-reorganization. Multi-point cross is implemented, you can set the cross probability threshold to ρc. The length of the chromosome is L, for random number 0 ≤ rj ≤ 1 (j = 1, 2, ..., L), if rj ≥ρc, then the next variable belongs to another gene, otherwise the next variable belongs to the previous variable A gene. Figure 1 Multi-point cross-example Fig.1 Multi Points Crossover optimal gene is the most suitable individual in the selection of genetically recombination and gene mutation. Gene mutation helps to jump out from local best, prevent premature convergence of algorithms. Set the mutant probability threshold is ρm, for random number 0 ≤ rj ≤ 1 (j = 1, 2, ..., L), if rj ≤ρm, then the jog in the chromosome has a mutation, otherwise, the original dyeing body is copied. Journal of J20. The specific implementation step of the gene algorithm is given below: (1) Generate random number, which constitutes the initial chromosome P0 = (A1, A2, ..., Al). Where AI is a chromosome, all parameters in the mathematical model consist of a particular arrangement. (2) Computational Function F (AI) of each chromosome. It is selected to be suitable for function F (AI) to sort, set the threshold, select the new parent chromosome P't. (3) Select the chromosome cross in a random manner. (4) During the biological evolution of nature, gene variability is very important. Due to the mutation of a certain two elements, the subgengene may have a greater adaptability than the parent gene. In the gene algorithm, the chromosome performs small unit variation in a relatively small mutation rate. (5) Calculate the adaptation function of each sub-gene, whether to achieve the condition, otherwise go to (2) to carry out the next step. 3 gene algorithm training CHMM
HMM is a generating model using a limited state system as a speech feature parameter, each state produces a continuous output feature. The HMM is actually a feature parameter generator that identifies voice based on the comparison of the visible voice parameters based on the parameters therebet. The decision based on the identification is the generation probability of the HMM model. In the process of introducing the gene algorithm into CHMM training, the first thing to solve is a structural problem of chromosomes. Arrange all the parameters of the CHMM model into a string, constitute a chromosome. For speech recognition, the HMM model from the left is used, which is 5 states only the CHMM model of one-stage jump from the left to right. The parameters in the CHMM model are composed of the initial state vectors π, the status transfer matrix A and the output probability density function of each state. The vector π contains 5 elements, in the A matrix, a common element 5 × 5 = 25, wherein the parameters of 0 are 12. More complicated is the output probability density function Bj (O) = σCJKN (O, μJk, UJK) of each state. Where: J represents the jth state; CJK is a mixed coefficient; n () is a Gaussian function; μJK is average vector; ujk is a covariance matrix. The voice signal is obtained by the microcomputer Sound Blaster Subcomaper, generating the WAV file sampling frequency is 11.025 kHz. The analysis frame length is 256 points, and each frame is shifted to 128 points. The 10-order LPC spectrum is used, and BJ (O) is selected into a mixture of 5 Gaussian probability density functions. The parameters of the vector π, the short array A and the mixing coefficient matrix C are formed into a string in line, form a front portion of the chromosome, the average vector μ and the covariance matrix U total 5 × 5 × ( 10 10 × 10) = 2 750 parameters Follow the latter part of the chromosome by line. In the CHMM model, the sum of the lines of the front portion of the chromosome is 1. It is also required to perform certain controls when producing a chromosome. When each generation of chromosomes are generated, each of the chromosomes corresponding to this part of the line vector can be normalized, and the constraint conditions of CHMM can be met. The Viterbi algorithm is used as an identification algorithm in the usual CHMM speech recognition. In other words, the observation sequence and the CHMM model are the most optimized target. Based on such thoughts, the suitable function of the gene algorithm is: all The observation sequence corresponding to the CHMM uses the Viterbi algorithm to see the sum of the observation probability, the larger the calculation results, the better the chromosome. In the previous portion of the chromosome in the experiment, two or more points are performed in probability, and then part of the chromosome only matches a multi-point cross, and the multi-point crossover is ρc = 0.8. Gene mutation probability ρm = 0.1 in front of the chromosome; for the latter part of the chromosome, ρM1 = 0.01, corresponding to gene mutation in units of parameters; ρM2 = 0.08, generated in units of rows. After the gene intersect or gene mutation, the normalization of the vector of the chromosome needs to be normalized. Each generation of genes is 300, and 60 excellent chromosomes are selected as a new parental gene, generate 240 chromosomes with genetically recombinant and genetic mutations, together with a new generation of chromosome. The training problem of the CHMM model has been converted to the problem of the maximum value of the observation sequence adaptation, and is solved by a gene algorithm. Table 1 Speech recognition system test results Tab.1Recognition Rate of Speech Recognition System0123456789NE1211312312NR49484949474948474948R (%) 989698989498969498964 Experimental results and discussion
