Research On Automatic Detection Of Patients With Mild Cognitive Impairment Based On Automatic Speech Recognition

As one of the advanced cognitive functions of mankind,speech and language are extremely complex processes,which involve extensive cognitive abilities.With the help of vocal organs,ideas are transformed into corresponding speech.At the same time,brain incorporates memory,knowledge,language and semantic information.Due to the decline of cognitive abilities,patients with mild cognitive impairment(MCI)experience difficulties in speech and language,with significant pathological changes in acoustics.This alteration in turn provides us with opportunities to detect early signs of MCI using speech technology.Most of the existing speech diagnostic system of MCI are built for Western language speakers,and models are trained with small datasets,which were collected using simple task paradigms.In addition,diagnosis in existing systems are based on human-defined features,which are oversimplified and fail to provide results with high precision.Considering the deficiencies of the existing systems,the present study aims to develop more sophisticated methods to automatically detect early signs of MCI using dataset of a greater size and that with multiple tasks in Chinese.Specifically,the present study includes the following three parts:1.Speech data of patients with MCI and healthy controls are collected.After pre-processing,a speech database of Chinese for the detection of MCI is established.The corpus contains 42.8 hours of speech recordings produced by 261 native speakers of Mandarin Chinese,which include 76 patients with MCI and 185 healthy controls.2.A high-precision automatic speech recognition(ASR)system is established based on speaker adaptation and domain adaptation,which compensates the hidden acoustic features and improves the performance of the ASR system.Speaker adaptation of feature-based Maximum Likelihood Linear Regression(fMLLR)and Learning Hidden Unit Contributions(LHUC)reduce character error rates(CER)onthe MCI dataset by 23.7% compared with the baseline system.A further reduction of CER by 15.1% was achieved when domain adaptation is applied.3.The automatic detection of MCI is studied in a comprehensive way from the perspective of not only ASR but also its underlying cognitive mechanisms.The present study achieves 0.75 and 0.82 of accuracy in speech fluency task and spontaneous speech production task,respectively.These results indicate that the methods proposed in the study can effectively detect patients with early signs of MCI.To summarize,the present study provides effective methods to detect early signs of MCI.