姜子星 叶武剑 吕友成 刘怡俊

摘要：病理语音是患者神经系统受损导致发音运动不协调所产生的语音。现有病理语音分析方法大多数采用频域语音特征如梅尔倒谱系数，并且识别模型也大多采用支持向量机模型。因此，提出一种时频特征短时傅里叶变换系数与梯度提升决策树的病理语音识别模型。首先，使用自建的脑卒中构音障碍数据集，提取语音的时频特征短时傅里叶变换系数。随后，结合梯度提升决策树算法进行分类识别。实验结果表明，提出的声学特征能够胜任脑卒中构音障碍识别任务。与梯度提升决策树分类器结合后，音节级别的准确率为68.5%，上升到说话人级别后准确率达到88.2%。

关键词：梯度提升决策树;构音障碍识别;时频特征

中图分类号：TN912.34      文献标识码：A

文章编号：1009-3044（2021）35-0131-03

Pathological Voice Recognition Based on Gradient Boosting Decision Tree

JIANG Zi-xing1， YE Wu-jian1， LV You-cheng2， LIU Yi-jun1

（Guangdong University of Technology， Guangzhou 510006， China; 2.Guangzhou Xinghai Integrated Circuit Center Co.， Ltd.， Guangzhou 510006， China）

Abstract： Pathological voice is the speech produced by uncoordinated pronunciation and movement caused by damage to the patient's nervous system. Most of the existing pathological voice analysis methods use frequency domain voice features such as Mel cepstrum coefficients， and most of the recognition models use support vector machine models. Therefore， proposed a pathological voice recognition model with time-frequency feature and gradient boosting decision tree. Firstly， using the self-built stroke dysarthria dataset to extract the short-time fourier transform coefficients of the time-frequency features of speech. Subsequently， the classification and recognition are carried out by the gradient boosting decision tree model. The experimental results show that the proposed feature can be competent for the recognition task of stroke dysarthria. After combining with the gradient boosting decision tree classifier， the accuracy rate of the syllable level is 68.5%， and the accuracy rate reaches 88.2% after rising to the speaker level.

Key words： gradient boosting decision tree; dysarthria recognition; time-frequency feature

1 引言

构音障碍是指由于患者中枢神经系统受损导致的发音运动不协调，从而导致患者发音混乱的现象，其严重程度决定于神经肌肉受损的程度[1]。研究调查发现，我国每年脑卒中新发病例超过250万人，已成为主要死亡原因之一[2]。因此能否有效诊断构音障碍对患者的预防及治疗起至关重要作用。

传统的构音障碍诊断方法主要是通过外科手术的方式。传统方法耗时且依赖于临床医生的主观判断，因此，研究人员和从业者一直在努力寻找这些手术的替代方法，基于语音样本的诊断就是其中之一[3]。语音信号的声学分析能够实现语音病理学的非入侵性、经济性、无偏见性和快速评估的优点[4]。……