用于连续语音识别的RBF-Gamma-HMM组合模型  

A RBF Gamma HMM Combined Model for Continuous Speech Recognition

作　　者：李易军[1] 徐近霈[2] 吴枫[2] 

机构地区：[1]北京大学计算机科学技术研究所栅格图象研究室,北京100871 [2]哈尔滨工业大学计算机系,哈尔滨150001

出　　处：《电子学报》1999年第9期81-85,共5页Acta Electronica Sinica

基　　金：国防科技预研基金

摘　　要：本文提供了一个有特色的、易扩展的多模块 Ｒ Ｂ Ｆ Ｇａｍ ｍａ 神经网与 Ｈ Ｍ Ｍ 组合的连续语音识别模型，兼有 Ｒ Ｂ Ｆ 网表达音元空间、 Ｇａｍ ｍａ 网综合时序相关信息、 Ｈ Ｍ Ｍ 作音元时间集成和扩展等功能，以实现功能互补．以本模型为基础，将本文提出的各种改进分类性能的学习算法用于特定人连续数字语音识别，其字正识率达到９８９ ％ ，串正识率达到９４８ ％ ．A continuous speech recognition model integrated by multi module RBF Gamma neural networks and HMMs is proposed in this paper.In this model,the abilities of RBF Gamma nets that effectively represent the space of speech units and synthesize the temporal correlation information of speech sequence are combined with the abilities of HMMs that integrate and expand the speech units in time domain.They are mutually complementary in function and improve the recognition accuracy obviously.A speaker dependent continuous digits recognition system is realized according to this model and using the learning algorithms proposed in this paper for improving classification performance.The tested digit accuracy is 98.9% and the string accuracy is 94.8%.

关 键 词：连续语音识别 Gamma神经网 隐马尔可夫模型 

分 类 号：TN912.34[电子电信—通信与信息系统]