Research On Task-Oriented Feature Representation For Speech Of Alzheimer’s Disease Patients

Speech-based recognition of Alzheimer’s disease has been proved to be an effective method.Compared with brain images and scales,speech is more economical and scalable,and can adapt to large-scale detection.Most of the current research methods are dimension-reduction-classification methods,which represent speech（usually more than 100,000 dimensions）as low-dimensional feature vectors（usually less than 100 dimensions）and then classify them to obtain disease recognition results.In feature representation,researchers try to use less restrictive CI（Content-independency）features to represent speech,but the traditional feature representation method lacks task orientation,resulting in a large number of task-related information lost in the process of dimensionality reduction.Based on task orientation,this paper uses metric learning method to improve the speech feature representation process and enhance the recognition accuracy of Alzheimer ’s disease on CI features.At the same time,through the interpretability framework,the feature representation method is explained in the model correlation and feature correlation.The main contributions of this paper are as follows:（1）This paper proposes a task-oriented speech feature representation method.By using the metric learning framework,the coding network can learn task-related information.The concept of dynamic boundary is introduced to improve triplet loss,and a measurement learning model based on DB-triplet loss is proposed to improve the accuracy of feature representation.On the data set of ADRess2020 Alzheimer’s disease speech classification challenge,the feature representation method in this paper exceeds all the acoustic feature representation methods in five models,and exceeds the linguistic feature representation method in some models,and obtains 83.3 %accuracy in the decision tree model.（2）In the early Alzheimer’s disease recognition task,this paper proposes a multi-objective task-oriented speech feature representation method.The paper takes the sample category learning as a goal,divides the tasks of multi-category learning into multiple single category learning goals,and trains multiple models on multiple goals to improve the focus of model learning.The polar coordinates are introduced to improve the unified metric loss function,and the Sphere loss is used to train the encoding network to help the model classify more accurately.This feature representation method surpasses the single-objective task-oriented method in experimental comparison,and obtains the best effect in the comparison experiment of similar features in the mixed feature representation method combined with linguistic features.（3）On the interpretability,the paper uses t-sne visualization model to visualize the features in the training process of the model,and finds that the spatial distribution of the features shows a clustering trend in the training process of the model.In terms of feature interpretation,this paper uses the SHAP framework to analyze the local and global information of features,and obtains the contribution details and importance ranking of features to the results.Based on this ranking,the paper uses the Grad-CAM framework to generate the corresponding thermal map for the more important features,and analyzes the deep color speech fragments in the thermal map to obtain the interpretation results of the corresponding features,which helps to better understand the model of the paper.