基于音节间相关识别单元的汉语连续语音识别算法

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 Ｎｂｅｓｔ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

Algorithm for mandarin continuous speech recognition based on context-dependent unit between syllables

In order to improve the performence of the recognition system for mandarin continuous speech, the context dependent units between syllables need to be constructed, and the recognition algorithm need to be studied. The theory of the mandarin continuous speech recognition based on DDBHMM and maximum a posterior estimation principle was discussed. And on the basis of this theory, in terms of the principles of dynamic programming, an algorithm for continuous speech recognition was proposed based on the context dependent unit between syllables. According to this algorithm, not only the optimal sentence candidate was obtained, but also the data structure of syllable lattice (that is the N best candidates) was obtained. At last, the error rate of the recognition system using the context dependent unit is reduced greatly in contrast to the system using the context independent unit.