一种基于改进CP网络与HMM相结合的混合音素识别方法

A Hybrid Approach for Phoneme Recognition Based on Combination of Improved CP Neural Network and HMM

提出了一种基于改进对偶传播 (CP)神经网络与隐马尔可夫模型 (HMM)相结合的混合音素识别方法。这一方法的特点是用一个具有有指导学习矢量量化 (L VQ)和动态节点分配等特性的改进的 CP网络生成离散 HMM音素识别系统中的码书。因此 ,用这一方法构造的混合音素识别系统中的码书实际上是一个由有指导 L VQ算法训练的具有很强分类能力的高性能分类器 ,这就意味着在用 HMM对语音信号进行建模之前 ,由码书产生的观测序列中已经包含了很强的分类信息 ,这将极大地改进 HMM系统在音素层上的识别性能。另一方面 ,由于这一训练是对一个具有诸多改进的 CP网络进行的 ,这就使得训练过程中的 LVQ学习能够自动地在有指导的方式下进行 ,而且加快了学习过程、改进了收敛性能、提高了分类精度 ,同时有效地减小了码书的大小 ,使得HMM的参数估计更为容易。最后 ,通过两个特定说话人的音素识别实验 ,将混合方法与使用 K -means聚类算法生成码书的 VQ- HMM传统音素识别方法进行了比较 ,实验结果表明混合系统的识别率能够达到 98%～ 99% ,误识率要比使用同样大小码书的 VQ- HMM识别系统的误识率低 4～ 6倍。

Proposes a hybrid approach for phoneme recognition based on combination of improved counter propagation (CP) neural network and hidden Markov model (HMM). The characteristic of the approach is that the codebook in a discrete HMM based phoneme recognition system is generated by a modified CP neural network with a few improvements, such as supervised learning vector quantization (LVQ) and dynamic node allocation. Hence, in effect, the codebook in the hybrid phoneme recognition system created through the approach is a high performance classifier with much better discriminating power trained by the supervised LVQ algorithm. It means that before a HMM is used for modeling a speech signal, the observation sequence generated by such a codebook contains highly discriminating information. This will greatly improve the recognition performance of HMM at phoneme level. On the other hand, since the training is done for an improved CP neural network with several new designs, the LVQ learning in the training process can be automatically performed in a supervised mode; learning is accelerated; system convergence is improved; more accurate classification is developed; at the same time, size of codebook is effectively reduced, resulting in the additional advantage of making HMM parameter estimation easier. Finally, through two speaker dependent phoneme recognition experiments, the hybrid approach is compared with the traditional VQ HMM phoneme recognition approach, which uses K means generated codebook. The results show that a correct recognition rate of 98%~99% can be achieved by the hybrid recognition system, and the error rate is 4~6 times lower than that of VQ HMM recognition system using a K means generated codebook of the same size.