Clustering of Voice Pathologies based on Sustained Voice Parameters

Alessa Anjos de Oliveira

1,2

, Maria E. Dajer

, Paula O. Fernandes

1,4 a

and João Paulo Teixeira

1,3,4 b

Polytechnic Institute of Bragança, Campus Sta. Apolónia, 5301 857 Bragança, Portugal

Federal University of Technology of Paraná, Campus Cornélio Procópio, 86300 000, Cornélio Procópio, Brazil

³Research Centre in Digitalization and Intelligent Robotics (CEDRI), Bragança 5300, Portugal

Applied Management Research Unit (UNIAG), Bragança 5300, Portugal

Keywords: Voice Pathologies Clustering, Clustering with Boxplot, Voice Pathologies Analysis, Jitter Shimmer HNR and

Autocorrelation Statistical Analysis.

Abstract: Signal processing techniques can be used to extract information that contribute to the detection of laryngeal

disorders. The goal of this paper is to perform a statistical analysis through the boxplot tool from 832 voice

signals of individuals with different laryngeal pathologies from the Saarbrücken Voice Database in order to

create relevant groups, making feasible an automatic identification of these dysfunctions. Jitter, Shimmer,

HNR, NHR and Autocorrelation features were compared between several groups of voice

pathologies/conditions, resulting in three identified clusters.

1 INTRODUCTION

Healthy individuals are able to produce vibrations in

the vocal folds periodically and with almost constant

intensity (Cordeiro, 2016). "When there are

pathological changes to the larynx, the level of signal

and its fundamental frequency change" (Panek et al,

2015), thus, this sound variation caused by a disorder

in the vocal tract may be audible or visible in certain

characteristics obtained by signal processing.

The assessment to detect vocal pathologies is

considered invasive and relatively expensive

(Cordeiro et al, 2015). In addition, the investigation

made by a specialist physician, as shown in Zwetsch

et al (2006), is not always accurate, because certain

laryngeal changes may be similar in certain prisms,

although they are intrinsically different from each

other.

An alternative method for the recognition of

laryngeal disorders is automatic detection using

speech processing, which, in reverse to the traditional

artifice, "enables non-invasive, low cost and objective

assessment of the presence disorders" (Panek et al,

2015).

Signal processing can be used to extract a set of

parameters that contribute to the detection of

https://orcid.org/0000-0001-8714-4901

https://orcid.org/0000-0002-6679-5702

laryngeal disorders. For this, several authors (Guedes

et al, 2018; Teixeira J. P. et al, 2017; Teixeira F. et al,

2018; Fernandes J. et al, 2019) used recorded signal

of sustained vowels from individuals with a healthy

voice, as well as, patients who have some voice

disorder. Thus, it is possible to extract relevant

information that serves to identify the pathologic

subjects or even to classify the pathology. For this

purpose, the following parameters have been used

extensively: absolute jitter, relative jitter, absolute

shimmer, relative shimmer, Harmonic-to-Noise Ratio

(HNR), Noise-to-Harmonic Ratio (NHR) and

autocorrelation.

Most studies using artificial intelligence to

identify the speech samples into one of the

pathologies report the scarcity number of subject

available for each class in the existent databases

(Teixeira F. et al, 2018). Anyhow, some databases

have a large number of pathologies available with a

low number of subjects. If this slightly small number

of individuals of any pathology could be grouped

with subjects of a statistically similar pathology, it

could become a larger group with a significant

number of subjects.

The present work aims to perform a statistical

analysis through the boxplot tool, based on the paper

written by Teixeira J. P. et al (2018, p. 172). It is used

280

Anjos de Oliveira, A., Dajer, M., Fernandes, P. and Teixeira, J.

Clustering of Voice Pathologies based on Sustained Voice Parameters.

DOI: 10.5220/0009146202800287

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 280-287

ISBN: 978-989-758-398-8; ISSN: 2184-4305

the previously quoted parameters taken from the

speech signal of 832 individuals for the possible

grouping of voice pathologies in order to create

relevant groups, making feasible an automatic

identification of these dysfunctions. The most

common symptoms that may indicate changes in the

larynx relate hoarseness, breathiness and roughness

(Teixeira J. P. et al, 2013, p 1112).

The section 2 of this document presents the

background of the boxplot statistical analysis,

followed with the description of material used in the

analysis in section 3. Section 4 describes the seven

parameters and for each one the statistical analysis

between the all pathologies. Finally, Section 5

presents the discussion and final conclusions together

2 DESCRIPTIVE STATISTICAL

ANALYSIS

The graphical presentation of a dataset made with the

boxplot or box-and-whisker plot uses five indicators:

the median, the first quartile, the third quartile, and

the smallest and largest number of the set (Mann,

2010). It helps to visualize the distribution and

skewness of the elements on the data set, in addition

to identifying outliers. To draft the box-and-whisker

plot is necessary to calculate the median, first

quartile, third quartile and interquartile range of these

elements (Hubert and Vandervieren, 2007). Then,

find the points that are 1.5 plus and minus this interval

in relation to the third and first quartile, respectively,

and, finally, the highest and lowest values of the set

are determined.

Figure 1: (a) First situation; (b) Second situation; (c) Third

situation.

According to Teixeira J. P. et al. (2018, p. 172) the

boxplot is used for a descriptive statistical analysis

and it compares the elements of a data set through

three situations: in the first situation there is no

overlap between the boxes, being B greater than A

(Figure 1(a)). Hence, there's a difference between

groups A and B. In the second situation (Figure 1(b))

the boxes overlay without their medians overlapping

the boxes, so there is, probably, a difference between

groups A and B. In the third situation (Figure 1(c))

the boxes overlap and their medians (at least one)

overlie the boxes. No difference can be considered

between groups A and B.

3 MATERIALS

The public available cured database described by

Fernandes et al (2019) was used in this work. This

database contains, among others, the seven

parameters used here, extracted by the algorithms

developed by Teixeira and Gonçalves (2016) and by

Fernandes et al (2018). This cured database was made

with the available features extracted from the

Saarbrücken Voice Database (SVD) (Barry and

Pützer). The SVD has, for each voice/subject, one

segment of voice record with the sustained vowels /a/,

/i/ and /u/ for High, Low and Neutral tones in a total

of nine speech segments. Each segment of the voice

consists in a steady state sustainable pronunciation of

the respective vowel. The individuals were analysed

and diagnosed by a physician. Many subjects

accumulated several voice pathologies, but for this

study, subjects with only one pathology were

considered. This was because the individuals with

several pathologies may interfere in the process of

segmentation of characteristics to identify the type of

dysfunction to be considered. Table 1 presents the

pathologies and its number of subjects available in the

cured database (Fernandes et al, 2019).

Table 1: Number of subjects for each pathology/condition.

Groups

Number of

Sub

ects

Carcinoma

Chronic Laryngitis

Control

Cyst

Functional Dysphonia

Granuloma

Hyperfunctional Dysphonia

Hypofunctional Dysphonia

Hypopharyngeal Tumor

Hypotonic Dysphonia

Intubation Granuloma

Laryngeal Tumor

Psychogenic Dysphonia

Reinke’s Edema

Spasmodic Dysphonia

Vocal Cord Paralysis

Vocal Cord Polyps

194

127

169

Clustering of Voice Pathologies based on Sustained Voice Parameters

281

For proper generalization, the sample size of each

group shall contain a relatively large number of

elements. Therefore, at least the pathologies Cyst,

Fibroma, Granuloma, Hypotonic Dysphonia,

Intubation Granuloma, Laryngeal Dysplasia are not

supposedly suited to any classification modelling

because they have a smaller sample size than the

‘relative large number of elements’. However, the

purpose of this work is to use pathologies with few

individuals in order to set them into relevant groups.

Thus, only fibroma and laryngeal dysplasia

dysfunctions will not be investigated because they

have only one individual.

4 PARAMETERS

CHARACTERIZATION AND

ANALYSIS

For this study, seven of the acoustic parameters

available in the cured database were used. The

parameters we considered were: absolute jitter,

relative jitter, absolute shimmer, relative shimmer,

HNR, NHR and autocorrelation.

4.1 Absolute Jitter

Absolute jitter, as pointed out by Teixeira J. P. et al.

(2018, p. 169), “is the glottal period variation between

cycles, that is, the mean absolute difference between

consecutive periods”. This description can be

presented in the form of an equation, as can be seen

in Equation 1.

𝑗

𝑖𝑡𝑡𝑎 =

𝑁−1



𝑇



−𝑇







(1)

The variable 𝑇



is the size of the glottal period and N

is the total number of glottal periods.

The comparison for absolute jitter was made

separating the genders as, in general, male voices

have low fundamental frequency and, consequently,

longer glottal periods. Therefore, for longer glottal

periods, there are larger deviations (Teixeira J. P. et

al 2018). Teixeira J. P. et al (2017) also highlights that

individuals, regardless of gender, have higher jitter

values when they cannot control the vibration of the

vocal folds.

4.1.1 Female Gender

In order to make an easier comparison Figure 2 shows

the boxplot of absolute jitter for all the pathologies

that are being worked on. In this framework, for cases

of hypopharyngeal tumor, hypotonic dysphonia and

granuloma caused by intubation, nothing can be

stated in the analyses, because there were not enough

female subjects in the database to print the boxplot.

Regarding all pathologies, a table can be created

with the samples and classify them according to the

boxplot presented in Figure 2, based on the previously

known theory shown in Figure 1. The number 1

points out the first situation (significant difference

between groups) as more appropriate, the number 2

presents the second situation (there is, probably, a

difference between groups) and the number 3

indicates the third situation (no difference between

the groups). Table 2 presents how were the analysis

made between each pathology with absolute jitter for

female gender in a succinct way.

4.1.2 Male Gender

In male-focused absolute jitter study, Table 3 shows

all pathologies together, pointing out the level of

intersection between them. Like in female gender

group, this table results from the statistical analyses

of all dysfunctions compared with themselves, like in

Figure 2.

4.2 Relative Jitter

According to Teixeira J. P. and Fernandes P. (2014),

this parameter “is the average absolute difference

between consecutive glottal periods divided by the

average period and expressed as a percentage”, as

shown in Equation 2.

𝑗

𝑖𝑡𝑡𝑒𝑟 =

𝑁−1

∑|

𝑇



−𝑇







𝑁

∑

𝑇







×100 (2)

The descriptive statistical analysis presented in

Table 4 succinctly presents a comparison between the

studied pathologies and their possible classification

based on how the dysfunctions are connected each

other, according to the boxplot overlap of each pair of

pathologies/conditions.

4.3 Absolute Shimmer

Absolute shimmer (ShdB) refers to the amplitude

variation peak-to-peak of a sound wave, in decibels,

extracted from a recorded signal that generates an

extensive and sustained vowel. This parameter can be

enunciated as the absolute mean of the multiplication

between a constant of value 20 and a base 10

logarithm of the ratio between two consecutive

periods, as can be seen in Equation 3.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

282

𝑆ℎ𝑑𝐵 =

𝑁−1

20×log

𝐴



𝐴









)

The variable 𝐴



is the magnitude of the glottal

period and N is the total number of glottal periods.

Based on the information provided in Table 5 it is

possible to compare the pathologies and classify them

according to this parameter.

4.4 Relative Shimmer

One of the parameters for voice acoustical analysis

that affects the vocal quality of patients is the relative

shimmer, which is determined, according to Teixeira

J. P. et al (2018, p. 170) "as the mean absolute

difference between magnitudes of consecutive

periods, divided by the mean magnitude, expressed as

a percentage", as presented in Eq. 4.

𝑆ℎ𝑖𝑚 =

𝑁−1

∑|

𝐴



−

𝐴







𝑁

∑

𝐴







×100 (4)

Table 6 shows the descriptive statistical analysis

of the relative shimmer parameter, which is used to

compare the pathologies each other based on the level

of intersection presented by their boxplot.

4.5 HNR

The HNR parameter represents the relationship

between the periodic and aperiodic components of a

speech segment, being the first component result

from the vibration of the vocal cords, while the

second component is a glottal noise. Several authors

inquire for different ways to express the HNR. In this

work, will be followed the Equation 5 presented by

Fernandes et al. (2018). According to the authors

HNR "consists in measure the energy of the first peak

of the normalized autocorrelation and consider that

this is the energy of the harmonic component of the

signal, and consider the remaining energy as the noise

energy given by the difference between 1 and the

harmonic energy”.

Table 7 resumes the comparison between

pathology groups and points out each one can be

considered the same group or not, based on the

existence of overlap between the boxplot’s groups,

like in Figure 2, and inform what is the level of

intersection between the groups.

𝐻𝑁𝑅(𝑑𝐵) = 10×log



𝑟





(𝜏



)

1−𝑟





(𝜏



)

(5)

The expression 𝑟





(𝜏



) is the maximum local

of the normalized autocorrelation.

4.6 NHR

Oppositely to the HNR, according to Fernandes

(2018), the NHR is the liaison between the aperiodic

component, more specifically the noise, and the

periodic component, related to the vibration of the

vocal cords. As this parameter is not measured in the

logarithmic domain, its values follow opposite

directions and are not exactly inverse to the HNR.

The Equation 6, which describes the NHR, is

determined as a function of the autocorrelation.

𝑁𝐻𝑅 = 1−𝑎𝑢𝑡𝑜𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛

(6)

The investigation of pathologies with results

presented in Table 8, were conceivable to analyze and

classify the dysfunctions among themselves, in order

to shortly express the relation between pathologies,

according to this parameter.

Figure 2: Comparison of boxplot of the pathologies groups using absolute jitter for female voice signals.

Clustering of Voice Pathologies based on Sustained Voice Parameters

283

Table 2: Result of the analysis of boxplot for absolute jitter for female voice signals.

Table 3: Result of the analysis of boxplot for absolute jitter for male voice signals.

Table 4: Result of the analysis of boxplot for relative jitter.

4.7 Autocorrelation

The correlation of two signals, by itself, is the sum of

the values of their products. Therefore, as Fernandes

(2018) said, the autocorrelation can be elucidated as

the correlation of a signal with itself. The

autocorrelation function, for a sound wave, is a

method of detecting the periodicity of the signal, in

which the autocorrelation of the signal window is

divided by the autocorrelation of the window used, as

can be seen in Equation 7.

𝑟



(

𝜏

)

𝑟



(𝜏)

𝑟



(𝜏)

(7)

where the expression 𝑟



(𝜏) is the normalized

autocorrelation of part of the selected signal and

𝑟



(𝜏) is the normalized autocorrelation of the used

window.

Table 9 presents the level of connection between

the pathologies.

Vocal Par Spasm D Reinke's Psych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D Granul Func D Cyst Chron La

Carcin Control

Vocal Pol

33333 32232322

Vocal Par

3333 33232322

Spasm D

333 3 33333 12

Reinke's

33 1 21223 21

Psych D

333233313

Laryn T

11131321

Intub Gran

Hypoph T

Hypot D

Hypof D

33333 13

Hyperf D

3333 13

Granul

33 2 1 3

Funct D

33 1 3

Cyst

313

Chron La

Carcin

Vocal ParSpasm D Reinke's Psych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

33331 3 3333133333

Vocal Par

333 2 3 3 33 31323 32

Spasm D

33 3 3 3 33 31323 32

Reinke's

31 3 3 3332333 23

Psych D

133333233333

Laryn T

11113111131

Intub Gran

33313333 33

Hypoph T

33 33333 23

Hypot D

333333 23

Hypof D

3333 3 23

Hyperf D

233 3 2 3

Granul

32 2 1 3

Funct

33 23

33 33

Cyst

323

Chron Lar

Carcin

Vocal Par Spasm D Reinke's Psych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

333323 3333232322

Vocal Par

3332 3 3 333232332

Spasm D

332 3 3 3332323 32

Reinke's

32 3 3 33 31323 32

Psych D

13 3333233323

Laryn T

11111111131

Intub Gran

3333333323

Hypoph T

33 32333 23

Hypot D

332333 13

Hypof D

32333 23

Hyperf D

233 3 2 3

Granul

23 2 1 3

Funct D

33 23

Cyst

313

Chron Lar

Carcin

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

284

Table 5: Result of the analysis of boxplot for absolute shimmer.

Table 6: Result of the analysis of boxplot for relative shimmer.

Table 7: Result of the analysis of boxplot for HNR.

Table 8: Result of the analysis of boxplot for NHR.

Vocal ParSpasm D Reinke's Psych D Laryn T Intub Gran Hypoph

Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

3 3331 3 3 333132332

Vocal Par

3332 3 3 333133333

Spasm D

332 3 3 333133333

Reinke's

32 3 3 323122332

Psych D

1 3 3 333233323

Laryn T

1 1 111111131

Intub Gran

3 333233333

Hypoph

333233323

Hypot D

33333323

Hypof D

32333 23

Hyperf D

233 3 23

Granul

22 2 1 2

Funct

33 23

33 33

Cyst

323

Chron Lar

Carcin

Vocal Par Spasm D Reinke's Psych D Laryn T Intub GranHypoph

Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

3 3 331 3 3 333132 3 32

Vocal Par

3 3 32 3 3 33 3133 3 33

Spasm D

332 3 3 333133 3 33

Reinke's

32 3 3 32 3122 3 32

Psych D

13 3333233323

Laryn T

11111111131

Intub Gran

3333233333

Hypoph

33 3233 3 23

Hypot D

33333 3 23

Hypof D

3233 3 23

Hyperf D

233 3 23

Granul

22 2 1 2

Funct

3323

3333

Cyst

323

Chron Lar

Carcin

ocal Par Spasm D Reinke'sPsych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

33331 3 3333233333

ocal Par

3 332 3 3 33 3233 3 33

Spasm D

331 3 3 33 3233 3 23

Reinke's

21 2 3 33 2122 3 32

Psych D

13 3333333323

Laryn T

11111111121

Intub Gran

3333333313

Hypoph T

33 3333 3 23

Hypot D

33333323

Hypof D

3333 3 23

Hyperf D

333 3 23

Granul

32 2 1 3

Funct

3323

Cyst

313

Chron Lar

Carcin

Vocal Par Spasm DReinke'sPsych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D Granul Func D Cyst Chron Lar Carcin Control

Vocal Pol

333313 3333233333

Vocal Par

3332 3 3 3332333 33

Spasm D

331 3 3 33 32333 33

Reinke's

31 3 3 33 31323 32

Psych D

13 3333333323

Laryn T

11111111121

Intub Gran

3333333323

Hypoph T

33 33333 23

Hypot D

333333 23

Hypof D

3333 3 23

Hyperf D

333 3 23

Granul

33 2 1 3

Funct D

33 23

Cyst

313

Chron Lar

Carcin

Clustering of Voice Pathologies based on Sustained Voice Parameters

285

Table 9: Result of the analysis of boxplot for autocorrelation.

Table 10: Results from the analysis considering the all parameters.

5 DISCUSSIONS AND

CONCLUSIONS

The descriptive statistical analysis of each criteria

presents the situation that best fits each comparison

of the sets. By observing each factor among all the

disturbances, it was possible to assemble Table 10,

which exposes whether or not the pathologies can be

grouped based on these parameters.

When comparing two elements, it can be observed

the values arranged for each of the seven parameters.

Thus, if a comparison has a value other than the third

situation (3) this analysis can already indicate that the

sets cannot be clustered. This is because, in order for

the sets to be joined, all the parameters must present

similarities for both pathologies. Hence, Table 10

presents the expressions "YES" and "NO" indicating

that the pathologies may or may not be gathered, as

well as the red color for the sets marked with the

second expression “NO”

As shown in Table 10, carcinoma, granuloma and

laryngeal tumor cannot be grouped with other

pathologies. Reinke’s edema, spasmodic dysphonia,

paralysis and polyps in the vocal cords can be

clustered together due to the similarity with each

other. Hyperfunctional, hypophunctional, hypotonic

dysphonia, hypopharyngeal tumor and intubation

granuloma can be bundled with functional dysphonia

cyst and chronic laryngitis.

The pathologies spasmodic dysphonia, paralysis

and polyps in vocal cords can agglomerate with

chronic laryngitis, granuloma by intubation,

hypopharyngeal tumor, functional, hypofunctional

and hypotonic dysphonias, as all factors can be

associated to the third situation.

It is visible that some pathologies like Chronic

Laryngitis, Hypotonic Dysphonia and

Hypopharyngeal Tumor can be grouped with all the

dysfunctions, except the groups that are not arranged

with any dysfunction.

Therefore, it is noticeable that the pathologies

with resemblance can be clustered, while divergent

dysfunctions are kept away from the other groups.

This work, in the future, can be used to confirm if

the use of Artificial Intelligence for clustering of

pathological diseases exhibit satisfactory results,

Vocal Par Spasm DReinke'

Psych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf DGranul Func D Cyst Chron Lar Carcin Control

Vocal Pol

3 3 331 3 3 33 32333 33

Vocal Par

3332 3 3 3332333 33

Spasm D

331 3 3 33 32333 33

Reinke'

31 3 3 33 32233 32

Psych D

13 3333333323

Laryn T

11111111121

Intub Gran

3333333323

Hypoph T

33 33333 23

Hypot D

333333 23

Hypof D

3333 3 23

Hyperf D

333 3 2 3

Granul

33 2 1 3

Funct D

33 23

Cyst

313

Chron Lar

Carcin

ocal Pa

Spasm D Reinke's Psych D Laryn T Intub Gran Hypoph T Hypot D Hypof D Hyperf D

ranu

Func D Cyst Chron Lar Carcin Control

Vocal Pol

YES YES YES YES NO YES YES YES YES NO NO YES NO YES NO NO

ocal Pa

YES YES YES NO YES YES YES YES YES NO YES NO YES NO NO

Spasm D

YES YES NO YES YES YES YES YES NO YES NO YES NO NO

Reinke's

NO NO NO YES Y ES NO NO NO NO NO Y ES NO NO

Psych D

NO YES YES YES YES YES NO YES YES YES NO YES

Laryn T

NO NO NO NO NO NO NO NO NO NO NO

Intub Gran

YES YES YES NO NO YES YES YES NO YES

Hypoph T

YES YES YES NO YES YES YES NO YES

Hypot D

YES YES NO YES YES YES NO YES

Hypof D

YES NO YES YES YES NO YES

Hyperf D

NO YES YES YES NO YES

Granu

NO NO NO NO NO

Funct D

YES YES NO YES

Cyst

YES NO YES

Chron Lar

NO NO

Carcin

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

286

since the sets of these pathologies, in this study,

already shows what pathologies can be clustered.

ACKNOWLEDGEMENTS

This work was supported by FCT – Fundação para a

Ciência e Tecnologia within the Projects:

UIDB/04752/2020 and UIDB/5757/2020.

REFERENCES

Barry, W.J., Pützer, M. Saarbrücken Voice Database,

Institute of Phonetics, Univ. of Saarland,

http://www.stimmdatenbank.coli.unisaarland.de/

Cordeiro, H. T., 2016. Reconhecimento de Patologias da

Voz usando Técnicas de Processamento da Fala.

Universidade Nova de Lisboa.

Cordeiro, H., Meneses, C., Fonseca, J., 2015. Continuous

Speech Classification Systems for Voice Pathologies

Identification. In IFIP AICT 450, pp. 217–224.

Fernandes

J., Silva, L., Teixeira, F., Guedes

V., Santos, J.

& Teixeira, J. P., 2019. Parameters for Vocal Acoustic

Analysis - Cured Database. In Procedia Computer

Science – Elsevier.

Fernandes, J., Teixeira, F., Guedes, V. Junior, A. &

Teixeira, J. P., 2018. Harmonic to Noise Ratio

Measurement - Selection of Window and Length. In

Procedia Computer Science - Elsevier. Volume 138,

Pages 280-285.

Guedes, V., Junior, A., Teixeira, F., Fernandes, J., &

Teixeira, J. P., 2018. Long Short Term Memory on

Chronic Laryngitis Classification. Procedia Computer

Science - Elsevier. Volume 138, Pages 250-257.

Hubert, M., Vandervieren, E., 2007. An Adjusted Boxplot

for Skewed Distributions. In Computational Statistics

& Data Analysis, Vol. 52.

Mann, P. S., 2010. Introductory Statistics, John Wiley &

Sons, Inc. Hoboken, 7th edition.

Panek, D., Skalski, A., Gajda, J., Tadeusiewicz, R., 2015.

Acoustic Analysis Assessment in Speech Pathology

Detection. In Int. J. Appl. Math. Comput, Vol. 25, No.

Shama, K., Krishna, A., Cholayya, N. U., 2006. Study of

Harmonics-to-Noise Ratio and Critical-Band Energy

Spectrum of Speech as Acoustic Indicators of

Laryngeal and Voice Pathology. In EURASIP Journal

on Advances in Signal Processing, Vol. 2007.

Teixeira, F., Fernandes, J., Guedes, V. Junior, A. &

Teixeira, J. P., 2018. Classification of

Control/Pathologic Subjects with Support Vector

Machines. Procedia Computer Science - Elsevier.

Volume 138, Pages 272-279.

Teixeira, J. P., Fernandes, J., Teixeira, F., Fernandes, P.,

2018. Acoustic Analysis of Chronic Laryngitis -

Statistical Analysis of Sustained Speech Parameters. In

Proceedings of the 11th International Joint Conference

on Biomedical Engineering Systems and Technologies,

pp 168-175.

Teixeira, J. P., Fernandes, P. O. & Alves, N. , 2017. “Vocal

Acoustic Analysis – Classification of Dysphonic

Voices with Artificial Neural Networks”, Procedia

Computer Science - Elsevier 121, 19–26.

Teixeira, J. P., Fernandes, P. O., 2014. Jitter, Shimmer and

HNR classification within gender, tones and vowels in

healthy voices. In Procedia Technology - Elsevier,

Volume 16, Pages 1228-1237.

Teixeira, J. P., Gonçalves, A., 2016. Algorithm for jitter and

shimmer measurement in pathologic voices. Procedia

Computer Science - Elsevier 100, pages 271 – 279.

Teixeira, J. P., Oliveira, C., Lopes, C., 2013. Vocal

Acoustic Analysis - Jitter, Shimmer and HNR

Parameters. In Procedia Technology – Elsevier, Vol. 9,

pp 1112-1122.

Zwetsch, I. C., Fagundes, R. D. R., Russomano, T., Scolari,

D., 2006. Processamento Digital de Sinais no

Diagnóstico Diferencial de Doenças Laríngeas

Benignas. In Scientia Medica, Vol. 16, No. 3.

Clustering of Voice Pathologies based on Sustained Voice Parameters

287