Psychotic Speech Recognition Method Based On Hybrid Feature Stacked Sparse Autoencoder

As relatively common psychotic,depression and schizophrenia have the characteristics of high prevalence,high suicide rate,and low treatment rate,which cause serious harm to individuals and society.Current diagnostic methods rely heavily on doctor’s clinical experience and patient self-description,and are greatly affected by subjective factors.Therefore,it needs an objective,effective and convenient evaluation method to detect psychotic.Among the many psychotic recognition modalities,speech has gradually attracted people’s attention because of its advantages such as no damage,low price,and rich emotional information.In the current research of many scholars,people’s views on the effective feature acquisition of speech diagnosis of psychotic are not consistent.In a small sample,how to obtain deep features that are highly complementary to existing features has not been well resolved.In solve to these problems,this paper conducts research on data collection,effective feature acquisition,and deep feature learning.The main contributions are as follows.(1)The speech data set of psychotic was collected and constructed.According to the results of clinician diagnosis,283 subjects were selected to divide the subjects into three classes: healthy people,depression and schizophrenia.The unified speech data collection was carried out.12 kinds of features were extracted according to the experience of previous people,and the data set of psychotic was constructed.(2)According to the characteristics of the data set constructed in this paper,the idea of sparse between feature groups is introduced.According to the contribution rate of different feature groups to psychotic recognition,a modified sparse Group Lasso feature extraction algorithm is designed to get the features that can effectively recognize psychotic.(3)Aiming at the shortcomings of the current original feature and deep feature fusion methods,this paper designs an hybrid feature stacked sparse autoencoder.This network fuses the original features into the training of the deep network,and can obtain more robust speech depth features.Based on this autoencoder,a psychotic speech recognition system was constructed.The system uses a feature selection algorithm based on L1 regularization to effectively remove the redundancy caused by the splicing of deep features and original features;and a w＿LDPP-SVM ensemble dimensionality reduction model is designed to further improve the generalization performance and stability.This paper constructs a three-classes psychotic speech data set of healthy people,depression,and schizophrenia for the first time.On this data set,we establish the speech recognition system for psychotic.The experimental results show that the speech recognition system of psychotic has better recognition accuracy than the current representative recognition method,which has certain guiding significance for speech recognition of psychotic.