论文部分内容阅读
针对目前嵌入式英语语音识别系统中识别性能较差或硬件资源占用较大的问题,提出了一个在16 b定点数据信号处理语音芯片上实现的非特定人、中等词汇量英语命令字识别系统。该系统采用基于连续隐含M arkov模型(con tinuous dens ity h idden M arkov m ode l,CDHMM)的两级识别网络,通过应用改进的音素体系、B ayes ian信息准则模型参数选择算法、决策树和数据驱动相结合的状态聚类方法、最小互信息改变准则特征选择算法,在保证识别率的前提下,大大降低了模型的存贮空间和计算复杂度。实验表明,对1 235词的英语短句的识别率为96.41%,识别时间为0.46倍实时。
Aiming at the problem of poor recognition performance or large occupation of hardware resources in embedded English speech recognition system, a non-specific and medium-vocabulary English command word recognition system implemented on 16 b fixed-point data signal processing speech chip is proposed. The system uses a two-level identification network based on the con tinuous M arkov model (CDHMM), through the application of improved phoneme system, B ayes ian information criterion model parameter selection algorithm, decision tree And the data-driven state clustering method, the minimum mutual information change criterion feature selection algorithm, under the premise of ensuring the recognition rate, greatly reducing the storage space and computational complexity of the model. Experiments show that the recognition rate of 1 235 English phrases is 96.41% and the recognition time is 0.46 times real-time.