Because of the excellent performance of Transformer in sequence learning tasks,such as natural language processing,an improved Transformer-like model is proposed that is suitable for speech emotion recognition tasks.T...Because of the excellent performance of Transformer in sequence learning tasks,such as natural language processing,an improved Transformer-like model is proposed that is suitable for speech emotion recognition tasks.To alleviate the prohibitive time consumption and memory footprint caused by softmax inside the multihead attention unit in Transformer,a new linear self-attention algorithm is proposed.The original exponential function is replaced by a Taylor series expansion formula.On the basis of the associative property of matrix products,the time and space complexity of softmax operation regarding the input's length is reduced from O(N2)to O(N),where N is the sequence length.Experimental results on the emotional corpora of two languages show that the proposed linear attention algorithm can achieve similar performance to the original scaled dot product attention,while the training time and memory cost are reduced by half.Furthermore,the improved model obtains more robust performance on speech emotion recognition compared with the original Transformer.展开更多
In order to improve the efficiency of speech emotion recognition across corpora,a speech emotion transfer learning method based on the deep sparse auto-encoder is proposed.The algorithm first reconstructs a small amou...In order to improve the efficiency of speech emotion recognition across corpora,a speech emotion transfer learning method based on the deep sparse auto-encoder is proposed.The algorithm first reconstructs a small amount of data in the target domain by training the deep sparse auto-encoder,so that the encoder can learn the low-dimensional structural representation of the target domain data.Then,the source domain data and the target domain data are coded by the trained deep sparse auto-encoder to obtain the reconstruction data of the low-dimensional structural representation close to the target domain.Finally,a part of the reconstructed tagged target domain data is mixed with the reconstructed source domain data to jointly train the classifier.This part of the target domain data is used to guide the source domain data.Experiments on the CASIA,SoutheastLab corpus show that the model recognition rate after a small amount of data transferred reached 89.2%and 72.4%on the DNN.Compared to the training results of the complete original corpus,it only decreased by 2%in the CASIA corpus,and only 3.4%in the SoutheastLab corpus.Experiments show that the algorithm can achieve the effect of labeling all data in the extreme case that the data set has only a small amount of data tagged.展开更多
基金The National Key Research and Development Program of China(No.2020YFC2004002,2020YFC2004003)the National Natural Science Foundation of China(No.61871213,61673108,61571106).
文摘Because of the excellent performance of Transformer in sequence learning tasks,such as natural language processing,an improved Transformer-like model is proposed that is suitable for speech emotion recognition tasks.To alleviate the prohibitive time consumption and memory footprint caused by softmax inside the multihead attention unit in Transformer,a new linear self-attention algorithm is proposed.The original exponential function is replaced by a Taylor series expansion formula.On the basis of the associative property of matrix products,the time and space complexity of softmax operation regarding the input's length is reduced from O(N2)to O(N),where N is the sequence length.Experimental results on the emotional corpora of two languages show that the proposed linear attention algorithm can achieve similar performance to the original scaled dot product attention,while the training time and memory cost are reduced by half.Furthermore,the improved model obtains more robust performance on speech emotion recognition compared with the original Transformer.
基金The National Natural Science Foundation of China(No.61871213,61673108,61571106)Six Talent Peaks Project in Jiangsu Province(No.2016-DZXX-023)
文摘In order to improve the efficiency of speech emotion recognition across corpora,a speech emotion transfer learning method based on the deep sparse auto-encoder is proposed.The algorithm first reconstructs a small amount of data in the target domain by training the deep sparse auto-encoder,so that the encoder can learn the low-dimensional structural representation of the target domain data.Then,the source domain data and the target domain data are coded by the trained deep sparse auto-encoder to obtain the reconstruction data of the low-dimensional structural representation close to the target domain.Finally,a part of the reconstructed tagged target domain data is mixed with the reconstructed source domain data to jointly train the classifier.This part of the target domain data is used to guide the source domain data.Experiments on the CASIA,SoutheastLab corpus show that the model recognition rate after a small amount of data transferred reached 89.2%and 72.4%on the DNN.Compared to the training results of the complete original corpus,it only decreased by 2%in the CASIA corpus,and only 3.4%in the SoutheastLab corpus.Experiments show that the algorithm can achieve the effect of labeling all data in the extreme case that the data set has only a small amount of data tagged.
