Voice Pathology Detection Using Deep Learning: a Preliminary Study

Abstract

This paper describes a preliminary investigation of Voice Pathology Detection using Deep Neural Networks (DNN). We used voice recordings of sustained vowel /a/ produced at normal pitch from German corpus Saarbruecken Voice Database (SVD). This corpus contains voice recordings and electroglottograph signals of more than 2 000 speakers. The idea behind this experiment is the use of convolutional layers in combination with recurrent Long-Short-Term-Memory (LSTM) layers on raw audio signal. Each recording was split into 64 ms Hamming windowed segments with 30 ms overlap. Our trained model achieved 71.36 % accuracy with 65.04 % sensitivity and 77.67 % specificity on 206 validation files and 68.08 % accuracy with 66.75 % sensitivity and 77.89 % specificity on 874 testing files. This is a promising result in favor of this approach because it is comparable to similar previously published experiment that used different methodology. Further investigation is needed to achieve the state-of-the-art results.

Changes in Phonation and Their Relations with Progress of Parkinson’s Disease

“…The advantage of this task in comparison with other commonly used vocal tasks is its independence of articulatory and other linguistic confounds [38]. Moreover, it is also present in most of the databases and therefore the experiments proposed in our work are comparable with other commonly used databases [39,40].…”

Towards robust voice pathology detection

“…For the above mentioned experiments, we decided to analyze the performance of the voice pathology detection models using multiple types of input data: a) raw audio samples to follow our previous work [22] and fur- ther explore possibilities of robust voice pathology detection without manually-selected features (DenseNet), b) conventional acoustic (dysphonic) features to follow the previously published works and quantify most common vocal pathologies (XGBoost, Isolation Forest), c) spectrograms to achieve a reasonable trade-off between dimensionality of the data and amount of information (DenseNet), and d) MFCC to follow the previous works focusing on voice and speech modelling, and voice pathology detection (all models).…”

“…livered state-of-the-art results in many domains including speech processing. To our best knowledge, despite our previous work [22], there are no other papers using deep learning algorithms for voice pathology detection. Next, we also employ the conventional voice pathology detection approach based on acoustic feature extraction procedure.…”

Convolutional Neural Networks for Pathological Voice Detection

“…Precision (P) shows how many of the pathological voice files classified are relevant, and F1-score (F1) has also been taken into account, calculated as in  It can be seen from Table V that the classifier achieved overall accuracy (ACC) of 88%, 66% and 77% on training dataset, validation dataset and testing dataset respectively. Compared to [11], spectrogram features show greater performance on pathological voice detection than raw timedomain signals. Moreover, the proposed algorithm is shown to be more robust for dealing with large amount of data compared to [6,8,9].…”

“…This is questionable compared to [10] using GMM-HMM which achieves 67.00% accuracy when the data amount is large. In [11], Deep Learning has been used for the first time, applying Long Short-Term Memory (LSTM), a type of recurrent neural network and using information from the timedomain axis. However, since pathological voice contains information without regard to time, this model might not be the most proper one for this problem.…”