论文部分内容阅读
为提高汉语连续语音识别系统的性能,建立了音节间相关的半音节识别单元,并研究了基于这种单元的连续语音识别算法。讨论了基于 D D B H M M 模型和最大后验概率估计准则的连续语音识别的理论基础,依据动态规划的基本原理,提出了一种基于音节间相关的识别单元的汉语连续语音识别算法。依照这种算法,不但能得到最优句子侯选,而且能够在识别过程中得到音节格(即 Nbest句子侯选)的数据结构。最后通过大词汇量非特定人连续语音识别的实验,表明了采用音节间相关的识别单元比基本的识别单元误识率有明显的降低
In order to improve the performance of Chinese continuous speech recognition system, the syllable recognition unit between syllables was established, and the continuous speech recognition algorithm based on this unit was studied. The theoretical basis of continuous speech recognition based on DABMH model and maximum a posteriori probability estimation criterion is discussed. Based on the principle of dynamic programming, a Chinese continuous speech recognition algorithm based on the unit of syllable correlation is proposed. According to this algorithm, not only can the optimal sentence candidate be obtained, but also the data structure of the syllable lattice (ie, N-candidate sentence candidate) can be obtained in the recognition process. At last, the experiment of non-specific continuous speech recognition by large vocabularies shows that the recognition rate between the syllable-related recognition units is lower than that of the basic recognition unit