关于连续语音识别率优化仿真研究

Research on Optimal Simulation of Continuous Speech Recognition Rate

连续语音识别是人机接口的关键技术,本质是双重随机可观测时变序列的音素参数识别过程,存在词汇量大、连续性强的技术难点。提出采用隐式马尔科夫模型来模仿过程,并借鉴Viterbi算法的解空间搜索点,对连续语音识别不存在单一HMM模型问题进行编码改进,使其适合于模型参数优化;同时,针对大词汇量情况下,存在模型参数非线性过强,对优化算法性能要求过高的问题,利用鱼骨结构对智能水滴算法进行改进,构建河道泥土量更新公式,扩大更新范围,从而提高搜索效率。最后,通过在标准测试函数和TIMIT语音数据库中的仿真对比,显示所提算法具有更快的运算速度和更高的预测精度,验证了算法的有效性。

Continuous speech recognition is the key technology of man-machine interface, the essence of the speech recognition is the process of the recognition of phoneme parameters of the time-varying sequence with double random, which has the technology of large vocabulary and strong continuity. Based on this characteristics, we used a hidden Markov model to simulate the process, and learned from the solution space point of Viterbi algorithm to solve the problem that single HMM model for continuous speech recognition does not exist, which makes it suitable to the model parametem optimization. Meanwhile, according to the problem of large vocabulary situation, there is strong nonlinear model parameter, which requires high performance of optimization algorithm, so, the bone structure was used to improve the intelligent water drops algorithm and the fiver mud weight updating formula was constructed to in- crease the update range, so as to improve the search efficiency. Finally, by comparing the simulation in the standard test functions with the TIM1T speech database, the result shows that the proposed algorithm has faster computing speed and higher prediction accuracy, which verifies the effectiveness of the algorithm.