摘要
利用隐马尔可夫模型(HMM)的动态时间序列建模能力及神经网络的模式分类能力,构成混合语音识别模型,同时考虑到语音信号的非平稳性,采用小波分析方法提取语音特征向量。通过时间规整方法,将所有具有可变长度的语音特征向量转换为相同维数的特征向量,从而简化了神经网络的结构。仿真结果表明,采用混合语音识别模型以及时间规整方法,不仅可提高识别率,同时大大缩减了训练时间,获得了很好的识别效果。
Applying dynamic time sequence modeling ability of hidden markov model(HMM) and classifying ability of neural networks, this paper presents a hybrid speech recognition model. Considering nonstationarity of phonetic signal, wavelet analysis method is used to extract feature vectors, which are conversed into fixed dimension with time alignment measure, as a result, the structure of neural network are simplified. Simulation results show that hybrid speech recognition model can result in both higher correct recognition rate and shorter training time.
出处
《电声技术》
北大核心
2004年第11期56-59,共4页
Audio Engineering
