Voice pathology detection using interlaced derivative pattern on glottal source excitation

Abstract

In this paper, we propose a voice pathology detection and classification method using an interlaced derivative pattern (IDP), which involves an n-th order directional derivative, on a spectro-temporal description of a glottal source excitation signal. It is shown previously that directional information is useful to detect pathologies due to its encoding ability along time, frequency, and timefrequency axes. The IDP, being an n-th order derivative, is capable of describing more information than a first order derivative pattern by combining all the directional information into one. In the IDP, first-order derivatives are calculated in four directions, and these derivatives are thresholded with the center value of each directional channel to produce the final IDP. A support vector machine is used as a classification technique. Experiments are conducted using three different databases, which are the Massachusetts Eye and Ear Infirmary database, Saarbrucken Voice Database, and Arabic Voice Pathology Database. Experimental results show that the IDP based features give higher accuracy than that using other related features in all the three databases. The accuracies using cross-databases are also high using the IDP features.

Healthcare Big Data Voice Pathology Assessment Framework

“…There are many feature extraction techniques involved in the VPA. These include Mel-frequency cepstral coefficients (MFCC) [23], multi-dimensional voice program (MDVP) [24], MPEG-7 low-level audio descriptors [25], IDP [26], and glottal noise parameters [27]. Each of them has its own advantages and disadvantages; however, after a careful consideration, we find that both the MPEG-7 audio features and the IDP features provided good results in the literature, and they are not too much affected by the diversity of the recorded signals.…”

“…They have been successfully applied to speech recognition [17] and voice pathology detection [26]. The IDP features have several interesting characteristics, which include a compact representation of second-order derivatives on four directions (time, frequency, time-frequency positive direction, and time-frequency negative direction), and a good encoding of timefrequency variations.…”

“…Take logarithms on the outputs of the filters. Step 4: Measure the IDP for each frame [26]. Figure 5 shows the time-domain signal and the IDP representation of a normal sample and a pathological sample.…”

A Deep Learning Method for Pathological Voice Detection Using Convolutional Deep Belief Networks

“…In Table 3, classification results in different metrics are listed. Sensitivity (SN) and Specificity (SF) are calculated in (6). Sensitivity reveals the performance on detecting the pathological voice files, and Specificity reveals the proportion of correctly detected healthy voice files.…”

“…However, healthy voice recordings and pathological voice recordings in this database are recorded in two different environments [6], which make it hard to distinguish whether it is discriminating environments or voice features. Saarbruecken Voice Database is a downloadable database with all recordings sampled at 50 kHz and with 16-bit resolution.…”