论文部分内容阅读
本文在指出隐马尔可夫模型（HMM）不合理假设的基础上,介绍了随机轨迹模型（STM）的理论机制及优越性。随机轨迹模型将语音基元的声学观察表示为参数空间中轨迹的聚类,并将轨迹建模为状态随机序列概率密度函数的混合,该模型可以克服HMM的不合理假设,在理论上更合理。根据STM的特点及汉语语音特色,本文对汉语连续语音识别基元的选取进行了讨论,提出了音素类单元作为识别系统的识别基元。基于STM的汉语连续语音识别的实验结果证明了STM的有效性和音素类单元的一致性。
Based on the unreasonable assumption of hidden Markov model (HMM), this paper introduces the theoretical mechanism and superiority of stochastic trajectory model (STM). The stochastic trajectory model expresses the acoustic observations of speech primitives as a cluster of trajectories in the parameter space and models the trajectory as a mixture of probability density functions of state random sequences. This model can overcome the irrational hypothesis of HMM and be more reasonable in theory . According to the characteristics of STM and Chinese phonetic features, this paper discusses the selection of Chinese continuous speech recognition primitives, and proposes the phoneme class unit as the recognition primitive of the recognition system. The experimental results of Chinese continuous speech recognition based on STM prove the validity of STM and the consistency of phoneme units.