一种改进的隐马尔可夫模型训练方法及其在声目标识别中的应用

摘要

In order to overcome the limitation of the classical HMM training algorithm and the generalization deficiency of existent discriminative learning method, a new discriminative training method for estimating CDHMM (continuous density hidden Markov model) in acoustic target recognition is proposed based on the principle of maximizing the minimum relative separation margin. By the definition of the relative margin, the new training criterion could be formulated as a standard constrained minimax optimization problem. Then the new HMM is got after the optimization problem is solved by a GPD (generalized probabilistic descent) algorithm. Experimental results prove that the new training criterion can achieve significant advancement over the former training method, which can improve the system performance effectively.%提出了一种基于最大相对界的改进隐马尔可夫模型训练方法.为解决隐马尔可夫模型的传统Baum_Welch训练算法在识别声目标时的局限以及现存区分训练算法泛化能力不足的问题,在经典隐马尔可夫模型为初始模型的基础上,定义了相对界,并通过最大化最小相对界建立一个最优化问题,用梯度下降法进行迭代求解,得到基于相对界的隐马尔可夫模型.将其应用于低空飞行目标声识别中,实际信号的识别结果表明算法相对于最好条件下的经典隐马尔可夫模型和最大互信息量隐马尔可夫模型性能有明显的提升,可以有效提高声目标识别系统的识别能力.