优化的竞争算法在语音识别中的应用

Speech Recognition Method Based on Competitive Learning Neural Networks

在传统的ＨＭＭ语音识别方法的基础上，提出了两种改进的竞争神经网络算法，分别用于语音识别的两个不同方面．首先提出了一种基于选择机制的新的竞争算法，这种算法可以有目的性地避免局部最优，而且可以克服模拟退火算法（ＳＡ）的随机性．然后，针对分类器的特性，对竞争算法进行改进，把安全拒识措施结合到竞争算法中，提出了一种新颖的神经网络——并行、自组织、层次神经网（ＰＳＨＮＮ）．实验结果表明，基于竞争神经网络算法的语音识别系统比传统的语音识别系统在识别能力和识别速度上都有明显提高，从而证明了与竞争神经网络算法结合的语音识别方法是可行的，而且具有良好的发展和应用前景．

Two new competitive learning approaches are presented for speech recognition in this paper. First, a new competitive learning algorithm is proposed with a selection mechanism, called the CLS (Competitive and Selective Learning) algorithm. As the selection mechanism enables the system to escape from local minima, the proposed algorithm can obtain better performance without a particular initialization procedure even when the imput data cluster in a number of regions in the input vector space. Next, a new neural network algorithm with competitive learning and multiple safe reflection schemes are proposed in the context of parallel, self organizing, hierarchical neural networks (PSHNN). The experimental results indicate that the recognition ability of the method based on competitive learning neural network is higher than that of the traditional HMM method, and prove the feasibility of the speech recognition method combined with competitive learning neural networks and its propective application.