基于免疫猫群优化算法的矢量量化的码书设计及语音识别

摘要

在矢量量化的码书设计过程中，针对传统的 LBG算法对初始码书选取的依赖性及易陷入局部最优的缺陷，提出基于免疫猫群优化算法的矢量量化码书设计。将整个种群分为搜索组和跟踪组，运用克隆扩增算子在搜寻组中进行局部搜索，根据适应度值大小调节变异个体数目，保持解的多样性。运用动态疫苗提取与接种算子使跟踪组个体基因与疫苗进行交叉变异，向最优解靠拢，防止无监督交叉变异可能引起的退化现象。通过浓度平衡算子和选择算子更新子代种群，防止种群“早熟”。将训练出全局最优码书输入到HMM模型进行训练和识别，实验结果表明，基于免疫猫群优化算法的矢量量化码书设计不依赖于初始码书选取，鲁棒性强且降低语音识别误差率。%In the process of codebook design, traditional LBG algorithm is often used for vector quantization which depends on the initial codebook selection and easily falls into local optimum. A vector quantization codebook design method based on immune cat swarm optimization algorithm ( ICSO) is proposed to solve the problems. The population is divided into searching group and tracking group. Clonal expansion operator is used for local search in the searching group, and the number of mutation individual is adjusted according to the fitness value. Moreover, dynamic vaccine extraction and vaccination operator are used for global search in the tracking group. The crossover and mutation between individual gene and vaccine make the result close to the optimal solution, and the descendant population is updated through the balance of concentration equilibrium operator and selection operator. Finally, the optimal codebook is obtained from the training vectors by the proposed algorithm and is inputted to the HMM model for training and recognition. The simulation results show that the proposed algorithm does not depend on the selection of initial codebook, has strong robustness and reduces the speech recognition error rate.