Quantitative Evaluation Of Elicitation And Analysis Of Feature Based On Depressed Speech

Clinical depression is a common mental disorder,which is characterized by prominent and persistent low mood.Constant depressive disorder can affect patients' individual physical and mental health.However,depression assessment methods make the diagnosis subjective,which can lead to high misdiagnosis rate.Many researchers search for a measurable indicator to assist clinicians to diagnose depression objectively.Among these indicators,speech is an attractive candidate for automated diagnosis systems,because it can be measured by non-invasive method easily and cheaply.Speech contains a wide variety of variability,one of these is speech elicitation.Among previous researches,there are some inconsistent results based on different speech elicitation.Therefore,we consider that speech elicitation is an important influence factor to acoustic feature.In this paper,we design an experiment involving speaking style and emotion stimuli to analyze the influence of speech elicitation through the statistical analysis and machine learning methods.In this paper,the main contributions and innovations are as follows:1.Through quantitative evaluation of speech elicitation,we find spontaneous speech and speech under neutral stimuli speech have high classification accuracies.Results are as following: 1)Compared to automatic speech,spontaneous speech performs more difference between the depressed and control with high classification rate.2)Compared to positive and negative emotion stimuli,speech under neutral stimuli can bring more difference between two groups,with high classification accuracy.2.Based on standardization of speech elicitation,the new features bring higher accuracy.Results are as following: 1)New feature performs better in difference between two groups.2)New feature brings higher classification accuracy,in detail,classification accuracy for female improves by an average of 5.4%,while male's average increases by 5%.3)standardization based on spontaneous speech and neutral stimuli perform better than other speech elicitation.3.After feature selection,the new feature set brings higher classification and efficiency.Results are as following: 1)Compared to the data before reducing dimension,classification accuracy for female improves by an average of 5.4%,while male's average increases by 5%.2)Maximum classification accuracy is 86.1% for female and 90.8% for male.In conclusion,depressed speech features are influenced by speech elicitation,in detail,spontaneous speech and speech under neutral stimuli perform better in significance testing and classification.The consequences of this research can provide a reference to optimize the experiment scheme in this research field.