基于混合DBNN-BLSTM模型的大词汇量连续语音识别  

Large vocabulary continuous speech recognition based on deep belief neural networks and bidirectional long-short term memory hybrid

作　　者：李云红[1] 王成 王延年[1] LI Yunhong;WANG Cheng;WANG Yannian(School of Electronics and Information,Xi′an Polytechnic University,Xi′an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《纺织高校基础科学学报》2018年第1期103-107,114,共6页Basic Sciences Journal of Textile Universities

基　　金：陕西省科技工业攻关项目(2016GY-047);陕西省科技厅自然科学基础研究重点项目(2016JZ026)

摘　　要：深度置信神经网络(DBNN)模型和双向长短时记忆神经网络模型(BLSTM)在单独进行特征提取时识别率不理想,长短时记忆单元(LSTM)与BLSTM模型可以更好解析语音数据特征.因此将DBNN模型和BLSTM模型相结合,提出一种大词汇量连续语音识别(LVCSR)的声学模型建立方法,并在Keras深度学习框架下进行实验.实验结果表明,使用改进的DBNNBLSTM模型进行大词汇量连续语音识别,识别精度有所提高,比BLSTM模型的语音识别率提高5%.The recognition rate is not ideal when the feature extraction is performed on the deep confidence neural network(DBNN)model and the bidirectional long-short term memory(BLSTM),the long-short term memory(LSTM)and BLSTM can better analyze the characteristics of speech data.By combining the DBNN model with BLSTM,a new acoustic modeling method for large vocabulary continuous speech recognition(LVCSR)is proposed and experimentally studied based on Keras deep learning framework.The experimental results show that the improved DBNN-BLSTM model has a high recognition accuracy,and the speech recognition rate is 5%higher than that of BLSTM.

关 键 词：大词汇量 语音识别 深度置信神经网络 双向长短时记忆模型 

分 类 号：TN912.34[电子电信—通信与信息系统]