摘要
语音情感识别的数据集普遍存在语音数据长短不一致的现象,进行补零处理会造成无用信息的冗余。针对此问题,提出了一种基于差分整合移动平均自回归ARIMA算法特征补齐的语音情感识别算法。首先进行特征的选取,并利用ARIMA方法进行语音特征的补齐。然后,基于因果扩张卷积神经网络和长短期记忆网络,构建语音情感识别模型。最后,采用柏林语音集进行实验,结果表明:用ARIMA方法对特征进行补齐,一定程度上提高了模型的表现力;使用因果扩张卷积搭建模型,增加了模型的泛用性。
The database of speech emotion recognition generally has the phenomenon of inconsistent length of data.The zeroize method will cause the superfluity of useless information.To deal with this problem,a speech emotion recognition algorithm based on feature complement by autoregressive integrated moving average(ARIMA)is introduced.Firstly,the features are selected and ARIMA is used to complement them.Then,a speech emotion recognition model based on causal expansion convolutional neural network and long shortterm memory network is constructed.Finally,experiments are carried out on the Berlin speech database.The results show that the features which are complemented by ARIMA improve the representability of the model to a certain extent and the model constructed by causal expansion convolution increases the versatility of the model.
作者
史少寒
周晓彦
李大鹏
SHI Shaohan;ZHOU Xiaoyan;LI Dapeng(School of Electronics and Information Engineering,Nanjing University of Information Science and Technology,Nanjing Jiangsu 211800,China)
出处
《电子器件》
CAS
北大核心
2023年第5期1333-1338,共6页
Chinese Journal of Electron Devices
基金
江苏高校品牌专业建设工程二期项目(电子信息工程)
教育部第二批新工科研究与实践项目(E-SXWLHXLX20202612)。
关键词
语音情感识别
差分整合移动平均自回归模型
长短期记忆网络
因果扩张卷积
特征补齐
speech emotion recognition
autoregressive integrated moving average
long short-term memory
causal expansion convolution
feature complement