全音节汉语语音识别系统的声学模型研究

A study on acoustic models of Chinese speech recognition system with whole Chinese syllables

本文在分析了汉语单音节发音的音节结构的基础上，定义了基于声韵母类的语音识别单元，从声学角度确定了用于汉语全音节识别系统的声学单元，并研究了这些声学单元的检测一致性及基于这些单元的识别系统的鲁棒性。文中还对大量的发音人的声母类发音的长度作了统计，给出了基于本文给出的分割算法和本文定义的声母单元下的声母长度预分类方法。通过在非特定人全音节汉语语音识别系统上的应用表明，本文定义的语音识别单元具有很高的检测一致性，建立在其上的识别，系统也具有很高的鲁棒性；文中给出的预选方法在最好的情况下，可以减少一半以上的运算量，而预选精度几乎达到１００％。

In this paper, we define initial-final-like units for Chinese speech recognition, based on the analysis of pronunciations of Chinese syllables, and determine the acoustic units used in the Chinese speech recognition system with whole syllables. We examine the detection consistence of these units and the robustness of the speech recoghtion systems constructed with these units. Through the applicahon of above units in a speaker-independent Chinese speech recognition system, it is shown that these units possess very good detection consistence, and the recognition system is very robust as for these units. In the paper, we count up, with a large number of utterances, the lengthes of initial-like units, which is derived from the segmentation method given in the paper. According to the recsults of the statistics, we present an algorithm of pre-selection. which can reduce up to half of the computation with almost 100 percent pre-selection accuracy.