text
stringlengths
0
512
Electrical and Computer Engineering, University of Southern California, Los
Angeles, California; ‡Keck School of Medicine, University of Southern California,
Los Angeles, California; and the §Department of Otolaryngology—Head & Neck
Surgery, University of Southern California, Los Angeles, California
Address correspondence and reprint requests to Benjamin van der Woerd,
Department of Surgery, Division of Otolaryngology—Head and Neck Surgery,
McMaster University, 50 Charlton Avenue East, Office G839, Hamilton, Ontario
L8N 1Y3, Canada
1Nikolaos Flemotomos is now working at Apple Inc
Downloaded for Anonymous User (n/a) at McMaster University from ClinicalKey.com by Elsevier on September 07,
2023
For personal use only
No other uses without permission
Copyright ©2023
Elsevier Inc
All rights reserved
within otolaryngology—head and neck surgery.15-17 With
respect to voice assessments, researchers have developed
tools to categorize samples according to gender and eval-
uate dysphonia using sustained vowel samples.15 However,
the traditional auditory-perceptual assessment includes
both sustained vowels and connected speech to evaluate
voice parameters across a variety of laryngeal beha -
viors.18,19 To date, there are no machine-learning global
assessments of dysphonia severity using both sustained
vowels and sentence samples calibrated to known expert
ratings
In this study, the authors seek to develop a ma-
chine-learning algorithm for evaluating dysphonia severity
on a 100-point scale using previously collected and expertly
rated voice samples of sustained vowels and connected
speech
METHODS
This study was designed using a previously labeled data set
to train a machine-learning model on
The Perceptual
Voice Qualities Database (PVQD) includes audio samples
(n = 295) which were professionally captured at partici -
pating voice centers
These samples include sustained vo-
wels and connected speech (CAPE-V sentences) segments
Furthermore, each sample was previously rated by three
experts on a 0–100 scale according to the standards of the
CAPE-V
The labeled data set allows the computer to
know how the samples are supposed to be rated and to fit
different criteria for the prediction model result in similar
estimates
The primary goal was to teach the model to
categorize voice samples according to the same scale
To do this, we used the OpenSMILE open-source toolkit
with the emobase2010 configuration.20 This was used to
extract acoustic and prosodic features, as well as pitch
onsets and recording duration
More explicitly, we ex-
tracted a base of 34 low-level descriptor (LLD) features
(including Mel-Frequency Cepstral Coefficient features,
logarithmic power of Mel-frequency bands, normalized
intensity, etc) with 34 corresponding delta coefficients ap-
pended and applied 21 different statistical functions (such
as standard deviation, arithmetic mean, skewness, kurtosis,
etc) to these, which resulted in 1428 features
Next, 19
functionals were applied to four LLD features based on the
pitch (F0final, jitterLocal, jitterDDP, and shimmerLocal)
as well as their corresponding four delta coefficients, re-
sulting in 152 features
Finally, we appended pitch onsets
and recording duration features
All LLD features are ex-
tracted based on a frame-by-frame analysis, using windows
of 25 ms with 10 ms frameshifts
In consideration of the nature and size of our dataset, we
adopted a support vector machine (SVM) instead of a re-
cursive neural network
An SVM is a method often
adopted for categorization problems, looking to draw lines
between categories that maximize the margin between the
line and the closest data points in each category
Within the
context of the SVM, different features were tried to achieve
the highest correlation and lowest root mean square error (RMSE) to the expert raters’ data
The features were
analyzed separately for different parts of the voice re-
cordings, and then combined to make the final predictions
Our feature set, extracted through OpenSmile, consists
of descriptive statistics related to low-level characteristics
These are widely utilized in various speech-related tasks
such as emotion recognition
The SVM model was im-
plemented with Gaussian kernel, and we tried different
configurations of the penalty parameter
C[10,1,10 ]1 1
and kernel coefficient
[10,10,10]4 3 2
The optimal
parameters were selected based on the five-fold cross-vali -
dation
This entails splitting the data set into five parts