语音理解中的容错技术的研究

本文研究了大词汇量非特定人汉语连续语音识别和理解系统中的容错技术.首先,声学识别器产生N个最优(N-best)音节候选及其相应的声学层的概念,再由N个最优音节候选构成一个音节网格(syllable lattice).一个容错语言分析器被用来搜索该音节网格并发现最优的汉字串.由于考虑了额外的可能候选音节,该最优汉字串的某些字的音节可能不在原来的音节网格中.这样,声学层的一些错误被纠正,语言分析器的稳健性(robustness)得以提高.实验表明容错分析器能将字的理解正确率从91.83%提高到94.15%.与传统的无容错技术的基于三元文法模型的分析器相比,错误率下降了28.4%.

A Study of Error-Tolerant Techniques for Speech Understanding

In this paper,error tolerant techniques are studied for large vocabulary speaker independent Chinese continuous speech recognition and understanding systems.When N best syllable candidates with their corresponding acoustic scores are generated and a syllable lattice is constructed by the acoustic recognizer,an error tolerant linguistic parser is used to search through the syllable lattice and find the optimal Chinese character sequence in which syllables of some characters may not be included in the lattice because additional possible candidates are considered.In such a way,some acoustic errors are corrected and the robustness of the parser is improved.Our experiments show that the error tolerant parser can increase the understanding rate from 91 83%,a 28 4% error reduction over the conventional trigram model based parser without error tolerant techniques.