摘要
在维吾尔语连续语音识别试验的声学层建模基础上,引用DDBHMM模型将上下文相关的三音子作为基本识别单元,并提出一种状态绑定的思想,对状态进行优化。为得到更充分的训练模型,提高识别效率,对语料库进行扩充,在多组对比试验的基础上,分析扩充前后对声学层识别速度、准确率等各个方面的影响。
DDBHMM(Duration Distribution Based HMM) is adopted as the acoustic model for Uyghur continuous speech recognition, and the context-dependent triphone model is selected as the best recognition unit, the Uyghur speech recognition system is optimised by using the state-binding method. In order to make the models be trained more sufficiently to improve the recognition performance, the corpus is enlarged, the emphasis is on analysis of the effect that the speech database's enlargement brings to the recognition rate and accuracy and so on based on several groups of contrasted experiments.
出处
《计算机工程》
CAS
CSCD
北大核心
2011年第2期197-199,共3页
Computer Engineering
基金
国家自然科学基金资助项目(60762006
60863008)
国家语委基金资助重点项目(MZ115-75)