语音听写机中语音、语言模型是两个非常重要的部分，而语音模型的好坏直接影响语言模型和听写机的性能。文中在一个大型数据库上对语音识别基元、语音模型、模型的输出观察向量的计分方法进行了大量的比较实验。实验表明，采取以音节为识别基元、基于中心距离正态分布的中心距离连续概率模型，和基于最近邻原则的输出观察向量计分方法即嵌入式多模板方案，可以取得很好的识别效果。 Speech dictation machine language, language model is two very important part, and the quality of the speech model directly affects the performance of the language model and dictation machine. In this paper, a large number of comparative experiments have been carried out on the scoring method of speech recognition primitive, speech model and model output observation vector in a large database. Experiments show that using syllable as recognition primitive, center distance continuous probability model based on center distance normal distribution, and output observation vector scoring method based on nearest neighbor principle, ie, embedded multi-template scheme, good recognition effect can be achieved .