A Transfer Learning Approach for the 2018 FEMH Voice Data Challenge

Human voice could be significantly affected by neoplasm, vocal palsy, and phono-trauma diseases. Computer aided diagnosis by analyzing human voice can be a remote and cost-effective tool for patients around the world. In this paper, we propose a deep transfer learning approach to differentiate pathological voice samples from normal ones. We utilize voice samples recorded from 200 patients at the Far Eastern Memorial Hospital (FEMH) to develop the deep transfer learning model. We extract prosodic, vocal tract and excitation features as new representations from the voice samples for diagnosis. To address the small data set challenge, we utilize the TIMIT dataset and develop a transfer learning approach in which a deep belief network (DBN) is first trained with the TIMIT data set. The trained model is then applied to the FEMH data set as a feature extractor. Finally, we train a support vector machine (SVM) classifier with the extracted features for diagnosis. We evaluate our approach using the leave one out cross validation (LOOCV) strategy on the 200 training patients, and achieve 94.90% sensitivity with 59.77% un-weighted average recall (UAR) for the 400 FEMH testing patients. Our results prove that the proposed method may be used effectively for pathological voice detection.

I. Introduction

The health condition of the human voice can significantly affect the quality of life of an individual [1]. The current practice to detect voice disorders accurately consists of a meticulous laryngeal examination procedure called "laryngeal endoscopy" [2], [3]. Performing this procedure is significantly expensive, as it requires physical intrusion in the patient’s body and an expert operator to perform the medical evaluation. A non-intrusive approach to detect voice disorder consists of analyzing the patient’s voice directly. Clinically, voice is considered as a multi-dimensional measurement tool and it can be used to identify the reason, severity, and prognosis of a disease, as well as to develop a therapeutic program for the cure [4]. Abnormalities in the respiratory system may lead to a voice problem and reduce mean airflow rate (MAFR). These phenomena can be further diagnosed to identify the clinical reason behind it [4], e.g.: vocal fold paralysis causes high MAFR whereas spastic dysphonia results lower MAFR [4]. Through patients phonating the vowels /a:/, /i:/ and /u:/, an expert can estimate the deviation from the optimum which can be useful for therapy modeling [4]. Although this method is non-intrusive, it still requires an experienced operator to give the diagnosis. In contrast, automated pathological voice diagnosis can be used as an inexpensive pre-screening tool to identify the disease remotely [2], [3].