摘要
对于目前在语音识别中广泛使用的两种技术即动态时间规整（DTW）技术和隐马尔可夫模型（HMM）的本质联系，提出了二者的统一模型（DHUM，DTWandHMMUni-fiedModel），并分别给出DTW和HMM向DHUM的转换关系。文中还提出了用DHUM解决更接近语音实际情况的高阶HMM作语音识别时所面临的运算量过大的问题。中等词表的识别实验结果表明，建立在DHUM之上的识别器的识别性能不低于DTW和HMM识别器。
Constructs a new DTW and HMM unified model (DHUM) by catching the essence connection between DTW and HMM. The transformations from DTW to DHUM and from HMM to DHUM are presented. DHUM seems to be a proper way to resolute the problem of high class HMM'S excessive calculation. The result of middle word corpus speech recognition (SR) test shows that the SR quality of DHUM is not lower than that of DTW and HMM.
出处
《数据采集与处理》
CSCD
1997年第3期218-222,共5页
Journal of Data Acquisition and Processing
基金
江苏省自然科学基金
南京理工大学科研发展基金
关键词
语音识别
动态时间规整
隐马尔可夫模型
speech recognition
speech processing
dynamic time warping
hidden Markov model
