【作者】 张家騄；

【Author】 Zhang Jialu Institute of Acoustics, Academia Sinica (Beijing P.O.Box 2712, 100080)

【机构】 中国科学院声学研究所；

【摘要】 本文简述根据语言可懂度理论所建立的汉语主观识别模型,列举了汉语在语音、语汇和语法上的主要特点,强调指出汉语声学模型和语言模型的特点在于:1．声韵调是适合汉语的音位系统;2．声学模型所运用的语音特征要首先区分发音方法进而区分发音部位;3．在音节层面上就要利用语言模型。本文还着重讨论了当前人们忽视的Bayes公式中分母项的作用,指出语言模型有助于提高识别系统的稳健性。文中还提出了分层链接语言模型,用于孤立单词和连续语言识别。更多还原

【Abstract】 The mathmatic models based on the speech intelligibility theory for Chinese syllables and words were described briefly and the characteristics of phonetic system, word constructions and grammer structures of Chinese were discussed in this paper. And then it is emphasized that: 1. The phoneme system of spoken Chinese is the initials, the finals and the tones; 2. The phonetic features uesd in acoustic model should be that it is easy to distinguish the manners of articulation at first and then the places of articulation; 3: The language model of Chinese should be used at from the syllable level. The role of the denominator of the Bayes famula. which is the general probability of the acoustic evidence, on the maximum likelihood estimation was disscused and it is shown that the language model is helpful to increase the robustness of automatic speech recognition systems. And a level-cascade language mode was poposed for both isolated word and continuous speech recognition systems foT Chinese.更多还原