论文部分内容阅读
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。
Aiming at the current low accuracy of HMM (hidden Markov model) widely used in continuous speech recognition, this paper proposes a recognition method of DDBHMM (duration distribution based HMM). According to the characteristics of Chinese, designed for continuous speech recognition triphone. Describes the MLSS (most likely statesequence) criterion used in recognition. The recognition network is designed and the frame synchronization identification algorithm for triphone recognition is illustrated. The comparison between the three-tone DDBHMM recognition method and the three-tone homogeneous HMM recognition method and the two-tone DDBHMM recognition method are compared. The results show that the recognition error rate can be reduced by 0.95% and 2.29% respectively by adopting the three-tone DDBHMM. This method can significantly improve the continuous speech recognition performance.