基于一维卷积神经网络和i-vector的孤立词语音识别  

Isolated Speech Recognition Based on One Dimensional Convolution Neural Networks and i-vector

作　　者：彭亘昌 姚干 周凯波 何顶新 Peng Genchang;Yao Gan;Zhou Kaibo;He Dingxin(School of Automation,Huazhong University of Science&Technology,Wuhan Hubei 430074,China)

机构地区：[1]华中科技大学自动化学院,湖北武汉430074

出　　处：《信息与电脑》2018年第4期43-46,共4页Information & Computer

摘　　要：针对孤立词语音识别的任务,采用一维卷积神经网络(CNN)作为声学模型,利用卷积核在时间轴上的移动来反映语音信号的时变性,从而提高识别性能。然后引入了反映说话人特征的身份认证矢量i-vector,通过i-vector特征消除说话人差异对识别造成的影响,进一步提高识别性能。在实验室自建语音库上的实验表明,基于这两种方法的识别系统达到了91%的识别准确率。For isolated speech recognition,a one dimensional(1D)convolution neural network(CNN)is used as the acoustic model.By convolution along the time axis,the temporal variation can be better satisfied and the recognition performance can be improved.Then,identity vector(i-vector)is introduced,which reflects the differences of acoustic characteristics between speakers and further improves the performance.Results on the homemade data set show that the combination of the two methods produces.91%accuracy rate.

关 键 词：语音识别 卷积神经网络 i-vector 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]