摘要
语音本身具有一定的上下文相关性,而传统语音识别系统中的语言模型对历史信息记忆能力不足,无法充分学习语音序列的相关性。为解决该问题,提出一种基于反向卷积的双向长短时记忆(Bi-LSTM)网络的语音识别方法,该模型在反向长短时记忆单元通路末端增加了一个卷积层,再经过两个全连接层,最后通过分类器输出识别结果。将该模型与目前主流的深度学习模型进行实验对比,结果表明该模型能有效提高语音识别正确率。
The speech itself has a certain degree of contextual relevance.However,the language model in the traditional speech recognition system is not capable of remembering historical information and can not sufficiently learn the relevance of the speech sequence.To solve this problem,this paper proposes a speech recognition method based on reverse convolutionary Bidirectional Long Short Term Memory（Bi-LSTM）network.The model adds a convolution layer to the end of the memory cell path in the reverse direction,and then passes through two fully connected layers.Finally,the recognition result is outputted through the classifier.Compared with the current mainstream depth learning model,this model can effectively improve the speech recognition accuracy.
作者
居治华
刘罡
陈琦岚
吕微
阮佳慧
武业皓
JU Zhi-hua;LIU Gang;CHEN Qi-lan;LV Wei;RUAN Jia-hui;WU Ye-hao(School of Computer Science,Hubei University of Technology,Wuhan 430068,China)
出处
《软件导刊》
2018年第7期27-30,36,共5页
Software Guide
基金
湖北省大学生创新创业训练计划项目(201710500032)
关键词
语音识别
双向长短时记忆神经网络
深度学习
speech recognition
bidirectional long short term memory neural network
depth learning
