Research On Speech Emotion Recognition And Its Application In Human-machine Dialogue System

With the gradual development of human-machine interaction,speech as a fast and direct way of communication,makes it more urgent for machines to understand emotions in the process of speech interaction.Due to language differences,different environments in which speech is recorded,uncertainties in the effective emotional features of speech,and noise in speech,emotional speech features lack of effective extraction and affect effectiveness in speech emotion recognition.Machine learning based speech emotion recognition methods only can extract specific features.Speech emotion recognition based on deep learning can avoid the above features extraction and automatically extract speech features.But these large number of speech features will reduce the model's recognition efficiency,and emotional speech features also cannot be effectively extracted.The purpose of this thesis aims to extract specific features and to get fusion features.Basing on deep learning algorithms,study a method of speech emotion recognition with strong generalization ability and high efficiency.And apply speech emotion recognition to human-machine dialogue systems.As speech emotion recognition,the effective extraction of emotional speech features is the key to the success of speech emotion recognition.This thesis proposes a complementary feature extraction method and verifies its effectiveness.In terms of corpora,corpus is an important element of the reliability of the verification method.This paper is based on two commonly corpus and a newer corpus for model training and verification.In terms of speech emotion recognition algorithms,efficient classification model algorithms determine whether the methods are effective.This thesis combines three speech emotion recognition corpus,as input of the deep learning recognition model.Our deep learning model is constructed by using an efficient convolutional neural network.The network structure has six convolutional layers and two fully connected layers.After each layer of convolutional layers,batch processing is used to normalize and adjust the sparsity of the network.During corpus training and verification of the model,ablation learning was performed on the complementary feature extraction method,and the corresponding results were compared with the three baseline methods.After experiments,the speech emotion recognition method proposed in this thesis is superior to the baseline methods.It proved the effectiveness of complementary feature extraction and the superiority of deep learning models.In this thesis,speech emotion recognition is applied to human-machine dialogue system,combining to construct a web dialogue system.The speech is used as the input of the system,and recognize the emotion of the speech.The corresponding emotion obtained is combined with the online speech recognition method.It proved that the speech emotion recognition method applied to the human-machine dialogue system successfully.