Autoencoder를 이용한 음성 신호 기반의 연하장애 검출 기법

Abstract

n this paper, a novel approach is proposed to improve speech-based dysphagia detection using an autoencoder. The water swallow test using voice is a convenient test method but has a problem due to low accuracy. Based on the /a/ voice data before and after the VFSS(Video Fluoroscopic Swallowing Study) test, which is a representative method for examining dysphagia, an autoencoder with a strong performance in detecting abnormalities is used to determine dysphagia. Data from 33 normal subjects were used for training. In 16 normal subjects and 39 patients with dysphagia, the proposed algorithm yields 23%p higher performance compared with the conventional Praat based scheme.