| As one of the basic ways of people’s communication,speech contains rich emotional information.Therefore,speech emotion recognition is of great significance for daily life and professional fields such as medical care and education.However,due to practical problems such as lack of laboratory equipment,lack of corpus data,and difficulty in emotional labeling,how to use limited corpus for efficient and fast training to improve the possibility of speech emotion recognition algorithms landing is still restricting the further development of speech emotion recognition.Therefore,in view of the above problems,this paper proposes a linear attention mechanism algorithm based on Taylor series,a transfer learning algorithm based on weighted gradient,and an emotion recognition model based on Transformer.The specific research contents are as follows:(1)The research background and application scenarios of speech emotion recognition are introduced.The research on the current state-of-the-art speech emotion recognition algorithms and excellent algorithms in other fields are described,as well as the development history and research status of the attention mechanism.At the same time,several common speech emotion databases are introduced.(2)Some preprocessing algorithms and common feature extraction algorithms for speech are introduced.The differences between feature sets are also analyzed and compared from the perspectives of extraction complexity,corpus emotion distribution,and emotion recognition performance,and are visualized through statistics and other methods.It proves the significance of different features for emotion discrimination.(3)The principles of various attention algorithms are introduced in detail from a mathematical point of view,and a linear attention mechanism based on Taylor series is innovatively proposed based on the correlation of matrix multiplication.Experiments are carried out on RNN,CNN and the proposed Transformer model using a variety of corpora.The performance of the proposed linear attention algorithm and that of Transformer model are compared from feature selection,emotion recognition performance and training performance.It is proved that the proposed linear attention algorithm not only has similar or better emotion recognition performance than the original one,but also can shorten the training time of the model to about half of the original one.At the same time,compared with the traditional RNN and CNN architecture models,the proposed Transformer model also has excellent recognition performance,and the UARs of CASIA,e NTERFACE and SUSAS are up to 5.28%,20.31%and 9.89% respectively.(4)In order to make full use of the limited training corpora,a transfer learning algorithm based on weighted gradient is proposed,which weights the gradient of each corpus in the process of back-propagation to reduce the impact of imbalanced data distribution.By introducing a gradient reversal layer to the corpus classification task,the differences of the corpus itself are further reduced,and the common representation of emotions is strengthened,thereby greatly improving the cross-corpus emotion recognition performance of the model.By designing multiple training strategies and conducting experiments on various corpora,the UAR of the proposed model,compared with that of the baseline and of the finetune model is improved by 10.19% and 5.52% on Emo DB,15.18% and 7.13% on ABC as well as 5.8% and4.63% on SUSAS respectively. |
