Research On The Method Of Detection Of Mind Cognitive Impairment Based On Speech Recognition

Mild cognitive impairment refers to the transition state between normal aging and Alzheimer’s disease,it’s reversible.However,Alzheimer’s disease is irreversible.Therefore,accurate screening of mild cognitive impairment can effectively prevent Alzheimer’s disease.As an advanced cognitive function,language can well reflect people’s cognitive ability.Therefore,for patients with mild cognitive impairment,when cognitive decline occurs,their language function will also be damaged,and it will also lead to changes in the pronunciation of their students,different from normal people.According to this characteristic,it can be analyzed that these voice changes by processing voice signals,so as to realize the screening of mild cognitive impairment.At present,some existing systems based on speech analysis to screen mild cognitive impairment were mainly in English or European languages.It has the problem of low matching for Chinese phonetic and grammatical features.And for Chinese,these kind of voice database also have the defects of small data scale,simple voice task and manual transcription,these will lead to lower efficiency and lower reliability of the system.Therefore,aiming at the above problems,we would study a screening method of mild cognitive impairment based on Mandarin speech,and design a practical and reliable mild cognitive impairment screening system.The specific work and research plan were as follows:(1)Establishing a Chinese mild cognitive impairment speech database.According to the goal of the project and the purpose of the experiment,we design data acquisition task,and the scheme of data acquisition has been established.And the voice from 74 MCI patients who have been diagnosed were collected,meanwhile voice from 122 normal old people were collected as controls in Luohu welfare center,Yuehai Street office,People’s Hospital of Shenzhen,and Tiantan Hospital of Beijing.At the same time,each user was assessed with MOCA scale,and the score of the assessment will be control the sample.The database of mild cognitive impairment in mandarin was established.A total of 196 samples were included in the database,and the total time of audio output from these samples is about 48 hours.(2)According to the speech characteristics of mild cognitive impairment patients,a filled pause detection technique has been presented.To improve the performance of ASR system,the technology of filled pauses detecting has been introduced into ASR.Classified several types of filled pauses without linguistic meaning in cognitive impairment speech by cascading machine learning model.Among them,the detection accuracy of filled pause in cognitive impairment speech was 91.7%,and the performance of ASR system was improved about 9% by introducing it into ASR system.(3)Multiple speech tasks were fused to screen patients with mild cognitive impairment.According to the features and functions of each kind of speech task,feature sets are established respectively.Through manual transcription and automatic transcription by ASR system,the feature sets of the three kinds of speech tasks are extracted and the feature vectors are constructed.By training the classification model,the discriminant probabilities of the samples on three kinds of tasks were obtained.Finally,the two kinds of methods obtained 0.94 and 0.85 screening accuracy respectively.On the level of a single sample,begin the correlation analysis between the results of the detecting system and the score of MOCA scale,it can prove the reliability of the detecting system.The experiments shown that the proposed method can effectively detect and screen mild cognitive impairment base on mandarin.