Research On Method And Key Technology For Depression Recognition Based On Speech

Depression is a kind of mental disorder which has symptoms like loss of interest and persistent feeling of sadness.It is the leading killer of mental illness with its high prevalence,high recurrence rate,high crippled rate and high fatality rate.The key to reduce the harm of depression is early diagnosis and early treatment,but one of the main obstacles is the lack of effective identification which is based on objective indicators.Through clinical observation and studies before,we can find out that the language behavior of depression patients have so many characteristics,such as slow,monotonous,low,pause.As a result,we believe depression diagnosis through speech signals is not impossible.And with its cheap and easy acquisition,the technology for depression recognition based on speech has become a new hotspot in recent years.Based on the research of depression recognition in speech,there are three key problems:(1)How to design an experiment to obtain high quality speech data.(2)How to select the effective features in numerous acoustic features.(3)How to construct efficient recognition model.This paper focused on these three issues,the main contributions are as follows:1.Speaking style and emotional valence are the main experimental factors in our experiment.Over the past two years,high-quality voice data of 536 participants were collected.It can be found that for depression recognition,interviews and picture descriptions are better than words reading,neutral valence is better than positive and negative one,this will be a promising choice of experimental paradigm and optimization provides reference model.2.In order to balance the three factors: operation time,classification accuracy,feature stability,we proposed an improved algorithm based on weighted amplitude FCBF.After verified its effectiveness,we also proposed a SHFS feature selection method,and verified its advantages on public data sets.Furthermore,we combined the results of several feature selection algorithms to select some of the effective features.3.On the basis of the preliminary determination of the effective features,we focused on the analysis of the features of pause,jitter,Mel cepstrum coefficient,linear prediction coefficient to filter the main role of the component.In order to measure depression speech monotonous,we proposed a new feature spectrogram based on entropy.The inspection found that depression patients and normal people have significant differences in most speech segments.In order to reduce the influence of individual differences,we proposed a feature normalization method for depression detection.Finally,based on the above results,we determined the effective speech feature set.4.In the study of depression recognition model,taking into account of the diversity of natural speech data,we put forward an ensemble learning model based on speech segments to select the optimal combination in multiple speech segments,which significantly improves the classification accuracy compared to the single voice segment model.At the same time,we also proposed an ensemble pruning method based on probability sample,it uses the sample sentence on the probability of choosing the base learner,and shows its superiority in public data.Finally,we optimized our ensemble learning model with this new pruning method.5.Based on the experimental design,effective features confirmation,model construction,we put forward the scheme of speech recognition system of depression and tested it on the data we have collected.The results showed that as long as the suitable ensemble model is selected,classification accuracy of male and female population can reach up to 79.8% and 70.1%.This paper focused on three key problems of depression recognition in speech.The solutions we proposed have a high accuracy.The data collected in this study and the proposed technical solutions have a positive effect on the use of speech as an objective indicator to identify depression quickly and effectively.