基于CDCPM的维吾尔语非特定人语音识别  

UIGHUR SPEAKER-INDEPENDENT SPEECH RECOGNITION BASED ON CDCPM

作　　者：王昆仑[1] 

机构地区：[1]新疆师范大学计算机科学系,乌鲁木齐830054

出　　处：《计算机研究与发展》2001年第10期1242-1245,共4页Journal of Computer Research and Development

基　　金：国家自然科学基金 ( 695 62 0 0 1);新疆维吾尔自治区"九五"重点攻关科研项目基金 ( G95 32 60 3)资助;荣获 1998年新疆维吾尔自治区人民政府"科学技术进步奖"三等奖

摘　　要：现代维吾尔语语音识别研究尚处于起始阶段 ,在此介绍了基于中心距离连续概率模型 ( CDCPM)的维吾尔语非特定人语音识别 .CDCPM用中心距离正态 ( CDN)分布描述模型特征空间 ,去掉了 HMM的状态转移概率矩阵 A,对 HMM进行了简化和改进 .在维吾尔语综合语音库上进行的实验表明 :恰当地估计模型状态数和模型混合密度数 ,当模型数为 5 2 5个 ,模型状态数为 16,混合密度数为 2 4 ,维吾尔语非特定人语音识别首选正识率达到97.90 % (集内 )和 94 .76% (集外 ) ,取得了较好的识别效果 .同时 ,指出了进一步开展维吾尔语语音识别研究的几个问题 .The Uighur speech recognition research is in the starting stage. Introduced in this paper is Uighur speaker independent speech recognition based on the center distance continuance probability model (CDCPM). CDCPM describes the feature space of model by center distance normal distribution (CDN), and simplifies and improves the HMM efficiently by getting rid of the state transition probability matrix A. A large amount of experimentation carried out with the Uighur synthetic speech database shows that numbers of state and amalgamate density for models can be adjusted adequately. When the number of model is 525, state number of model is 16, and the amalgamate density number of model is 24, the rate of first correct recognition is up to 94 76% (in set) and 97.90%(out set) on the Uighur speaker independent speech recognition. Recognition result with good performance is derived. At the same time, some problems about Uighur speech recognition research are pointed out.

关 键 词：维吾尔语 语音识别 中心距离连续概率模型 CDCPM 

分 类 号：TN912.34[电子电信—通信与信息系统]