Analysis Of Infant Crying Sentiment Based On Deep Learning

Sentiment analysis is one of the main research directions of natural language processing in artificial intelligence,which has attracted much attention from domestic and foreign research institutions and related personnel in recent years.Sentiment analysis includes text oriented,picture oriented,video oriented,speech-oriented and other fields.As speech is the most primitive and direct way of human communication,which contains rich emotions of human beings,emotion analysis of speech is particularly important.The purpose of this paper is to find an intelligent model which can automatically recognize the emotion of baby cry.To achieve the baby's physical health,psychological development and mental state to make an accurate judgment,so as to effectively help the healthy growth of the baby.That was the reason why he was so much more important for the search that he made for the application.On the basis of previous research,this paper mainly does the following work in order to solve the problems that need to be solved in the analysis of infant crying sentiment:(1)In view of the fact that there is no standard,standardized infant crying database that has caused the researchers to classify infant crying categories,this article has consulted a large number of documents,and conducted analysis and research on this basis.According to Dunstan baby language theory,a baby crying database containing five emotions was designed and established,which provides a reliable data set for subsequent model training and other researchers in the field.(2)Aiming at the pre-processing of baby crying data,this paper studies the characteristics of baby crying voice,summarizes the differences and characteristics of baby crying and adult voice,and proposes the use of pre-emphasis,framing,and windowing based on the characteristics of baby crying data.,The endpoint detection method performs crying data preprocessing to provide high-quality audio data for subsequent generation of spectrograms.(3)In order to improve the efficiency of sentiment recognition rate and model training time of existing infant cry emotion classification model,a Convolutional Neural network(CNN)was proposed.and Transformer model and other deep learn-based speech sentiment analysis technology advantages and shortcomings of Long-Short Term Memory(LSTM).Then this paper proposes to classify infant crying sentiment based on CNN-Transformer model.The model extracts the feature image of the language spectrum by using CNN,compressing the extracted feature image into vectors,then inputs the vector into the encoder part of transformer model to model the continuous crying of infants.Finally,CNN-Transformer model is improved by adopting modified position coding,multi-head sensing,residual linking and normalization,and fully connected feedforward network.(4)In order to accurately evaluate the proposed model,this paper trains the proposed CNN-Transformer infant crying emotion analysis model on the established database.In order to prove the fairness of the experiment,the public ESC-50 data set and Freesound data are used.The experimental results show that the model is improved obviously in the classification accuracy rate,single classification accuracy rate,and training time consumption.To sum up,this paper fully analyzes and studies the characteristics of baby cry,designs a baby cry database containing five emotions,and proposes a cry emotion analysis model based on CNN-Transformer,which is verified and analyzed by experiments.The work done in this paper provides a new research idea and method for other fields of speech sentiment analysis.