基于汉语语音音位的HMM建模方法

为了减少声学模型复杂度、降低对嵌入式系统的硬件资源需求,提出了为汉语全音节的声母、韵首、韵腹、韵尾4部分音位分别建立隐含Markov模型的新方法。基于汉语语音学的音位知识,并结合4部分音位方案比较实验,最终确定声母、韵首、韵腹、韵尾4部分音位模型总数分别为76、12、76、14,对应的4部分的模型状态数分别为4、1、4、2。同采用声母、韵母2部分建立的半音节隐含M arkov模型相比,新系统中模型数、状态数减少了30.2%、36.5%,同时关键词识别率提高1.32%。

HMM modeling based on mandarin phonemes in embedded systems

A method of acoustic model design was developed for Hidden Markov Models to reduce the complexity of the acoustic models and lower the hardware requirements in embedded systems.The method separately models each initial,glide,nucleus,and coda phoneme.The model numbers of these four parts were 76,12,76,and 14,and the state numbers of each model of these four parts were 4,1,4,and 2 in the final system based on the knowledge of mandarin phonemes and the results of scheme comparison tests.The total number of models was reduced by 30.2% with the number of states was reduced by 36.5%.The key word detection accuracy was improved by 1.32% compared with the method of modeling each initial and final semi-syllable.