Voice Pathology Detection and Classification by Adopting Online Sequential Extreme Learning Machine

Fahad Taha Al-Dhief,Mazin Abed Mohammed,Musatafa Abbas Abbood Albader,Nurul Mu'Azzah Abdul Latiff,Nik Noordini Nik Abd Malik,Marina Mat Baki,Nor Muzlifah Mahyuddin,Naseer Sabri Salim

doi:10.1109/access.2021.3082565

Abstract

In the last decade, the implementation of machine learning algorithms in the analysis of voice disorder is paramount in order to provide a non-invasive voice pathology detection by only using audio signal. In spite of that, most recent systems of voice pathology work on a limited acoustic database. In other words, the systems use one vowel, such as /a/, and ignore sentences and other vowels when analyzing the audio signal. Other key issues that should be considered in the systems are accuracy and time consumption of an algorithm. Online Sequential Extreme Learning Machine (OSELM) is one of the machine learning algorithms that can be regarded as a rapid and accurate algorithm in the classification process. Therefore, this paper presents a voice pathology detection and classification system by using OSELM algorithm as a classifier, and Mel-frequency cepstral coefficient (MFCC) as a featured extraction. In this work, the voice samples were taken from the Saarbrücken voice database (SVD). This system involves two parts of the database; the first part includes all voices in SVD with sentences and vowels /a/, /i/, and /u/, which are uttered in high, low, and normal pitches; and the second part utilizes voice samples of the common three types of pathologies (cyst, polyp, and paralysis) based on the vowel /a/ that is produced in normal pitch. The experimental results have shown that OSELM was able to achieve the highest accuracy up to 91.17%, 94% of precision, and 91% of recall. Furthermore, OSELM obtained 87%, 87.55%, and 97.67% for f-measure, G-mean, and specificity, respectively. The proposed system also presents a high ability to achieve detection and classification results in real-time clinical applications.

Highlights

Voice pathology analysis is considered a very significant field in the healthcare area
This paper is organized as follows; Section II discusses the drawbacks of the state-of-the-art in the systems of voice pathology detection and classification; Section III describes the proposed methods in terms of Saarbrücken voice database (SVD), Mel-frequency cepstral coefficient (MFCC), and Online Sequential Extreme Learning Machine (OSELM) classifier; Section IV presents the experimental results of OSELM for all SVD in general and for three common pathologies in particular
THE PROPOSED METHODOLOGY In this work, we propose a voice pathology detection and classification system using OSELM technique in which the algorithm has been demonstrated to be extremely fast with good generalization performance [37]

Summary

INTRODUCTION

Voice pathology analysis is considered a very significant field in the healthcare area. The study of speech signal processing of pathological voices by objective evaluation becomes an important topic for researchers as it aims to reduce medical laboratory work in diagnosing pathological speeches It provides a non-invasive method of diagnosis which is more comfortable for patients, faster, as well as cost-effective [11]. Machine learning algorithms can serve as objective evaluation tools for speech processing in order to detect pathological voices from acoustic recordings. This paper is organized as follows; Section II discusses the drawbacks of the state-of-the-art in the systems of voice pathology detection and classification; Section III describes the proposed methods in terms of SVD, MFCC, and OSELM classifier; Section IV presents the experimental results of OSELM for all SVD in general and for three common pathologies in particular.

RELATED WORK

EXPERIMENTAL RESULTS AND DISCUSSION

CONCLUSION AND FUTURE WORK