论文部分内容阅读
在基于特征的语音识别研究中,往往会发现其中有些特征的识别性能对一些音比另一些音更好,而另一些特征却与此相反。它们在一些音的识别特性上存在着一定程度的互补。本文基于目前话音识别研究主要方法之一的HMMM识别方法,提出了三种有效综合利用这种互补关系提高HMM识别性能的方法。作者分别称它们为顶尖参数法,全部参数法和最可靠参数法。这三种方法在多发音人汉语数字的DHMM/VQ语音识别中,分别将识别率由89％提高到了92.3％、95.7％、94.3％。本文将详细介绍这三种方法,及其在多发育人汉语数字的DHMM/VQ语音识别中试验结果极及其分析。
In the research of feature-based speech recognition, it is often found that some of these features have better recognition performance for some than others, while others are the opposite. There is some degree of complementarity in the recognition of some sounds. In this paper, based on the HMMM identification method, one of the main methods of speech recognition, this paper proposes three methods to effectively improve the recognition performance of HMM by comprehensively utilizing this complementary relationship. The authors call them the top parameter method, the full parameter method and the most reliable parameter method, respectively. In the DHMM / VQ speech recognition of multi-speaker Chinese numerals, the three methods respectively increase the recognition rate from 89% to 92.3%, 95.7% and 94.3%. This article will detail the three methods and their test results in DHMM / VQ speech recognition for multi-developmental Chinese speakers.