论文部分内容阅读
为提高语音识别系统的性能,针对汉语语音的单音节结构的特点,提出了建立三音子识别单元的方法。这种方法完全利用语音学知识对上下文进行分类从而实现参数共享,而不同于传统的数据驱动的聚类共享。提出并实现了采用三音子单元的识别系统的训练算法和识别搜索算法。实验表明:基于语音学分类的三音子单元对识别性能有明显的改善,系统的首选误识率相对基线系统降低了28%。
In order to improve the performance of speech recognition system, aiming at the characteristics of monosyllable structure of Chinese speech, a method of establishing triphone recognition unit is proposed. This method completely uses phonetics knowledge to classify the context to achieve parameter sharing, as opposed to the traditional data-driven clustering sharing. The training algorithm and recognition algorithm of recognition system using triphone unit are proposed and implemented. Experiments show that the triphone unit based on phonetic classification has obvious improvement on the recognition performance, and the system’s preferred misclassification rate is reduced by 28% compared with the baseline system.