【摘要】 针对目前连续语音识别中广泛使用的齐次HMM(hidden Markov model)模型识别精度低的现状,该文提出了三音子DDBHMM(duration distribution based HMM)识别方法。根据汉语的特点,设计了适用于连续语音识别的三音子。描述了识别中使用的MLSS(most likely statesequence)准则。设计了识别网络并阐明了用于三音子识别的帧同步识别算法。将三音子DDBHMM识别方法与三音子齐次HMM识别方法和双音子DDBHMM识别方法进行了实验对比,结果表明:采用三音子DDBHMM可以使得识别错误率分别下降0.95%和2.29%。说明该方法能够显著地改进连续语音识别性能。更多还原

【Abstract】 The HMM(hidden Markov model) is widely used in continuous speech recognition,but the recognition rate can still be improved.This paper presents a triphone DDBHMM(duration distribution-based HMM) recognition method to improve the recognition performance.The triphone used for the continuous speech recognition was designed based on the Chinese language characteristics.MLSS(most likely state sequence) rules are described with a recognition network designed based on the frame-synchronous recognition algorithm.The triphone DDBHMM recognition method gives a 0.91% better recognition rate than the triphone HMM recognition method and 2.29% better rate than the diphone DDBHMM recognition method.Thus,this method significantly improves the recognition performance of continuous speech recognition.更多还原