论文部分内容阅读
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 (syllablelattice) .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 1 5 % .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 2 8 4% .
In this paper, we study the fault tolerant techniques in the Chinese system of continuous speech recognition and understanding for large vocabulary non-specific speakers.Firstly, the concept of N best syllable candidates and their corresponding acoustic layers is generated by the acoustic recognizer, The optimal syllable candidates form a syllable lattice. A fault-tolerant language analyzer is used to search the syllable lattice and find the optimal Chinese string. Since the extra possible candidate syllables are considered, The syllables of these words may not be in the original syllable grid, so that some errors in the acoustical layer are corrected and the robustness of the language analyzer is improved.Experiments show that the fault-tolerant analyzer can improve the understanding of words from 91 83 % To 94 15%. Compared with the traditional ternary grammar-based analyzer without fault tolerance, the error rate dropped by 284%.