Applied Computer Science, vol. 19, no. 4, pp.
14–25
doi: 10.35784/acs-2023-32
Submitted: 2023-09-05 | Revised: 2023-10-19 | Accepted: 2023-11-05
Keywords: Dysphonia, Machine learning, Multiclass classification, Voice signal
Miguel Angel BELLO-RIVERA [0009-0003-6641-3094]*,
Carlos Alberto REYES-GARCÍA [0000-0003-4773-9585]**,
Tania Cristal TALAVERA-ROJAS [0000-0001-7656-3115]***,
Perfecto Malaquías QUINTERO-FLORES [0000-0001-7651-4364]*,
Rodolfo Eleazar PÉREZ-LOAIZA [0000-0002-6500-258X]*
AUTOMATIC IDENTIFICATION OF DYSPHONIAS
USING MACHINE LEARNING ALGORITHMS
Abstract
Dysphonia is a prevalent symptom of some respiratory diseases that affects voice
quality, even for prolonged periods. For its diagnosis, speech-language pathologists
make use of different acoustic parameters to perform objective evaluations on patients
and determine the type of dysphonia that affects them, such as hyperfunctional and
hypofunctional dysphonia, which is important because each type requires a different
treatment. In the field of artificial intelligence this problem has been addressed through
the use of acoustic parameters that are used as input data to train machine learning
and deep learning models. However, its purpose is usually to identify whether a patient
is ill or not, making binary classifications between healthy voices and voices with
dysphonia, but not between dysphonias. In this paper, harmonic-to-noise ratio, cepstral
peak prominence-smoothed, zero crossing rate and the means of the Mel frequency
cepstral coefficients (2-19) are used to make multiclass classification of voices with
euphony, hyperfunction and hypofunction by means of six machine learning algorithms,
which are: Random Forest, K nearest neighbors, Logistic regression, Decision trees,
Support vector machines and Naive Bayes. In order to evaluate which of them presents
a better performance to identify the three voice classes, bootstrap.632 was used. It is
concluded that the best confidence interval ranges from 87% to 92%, in terms of
accuracy for the K Nearest Neighbors model. Results can be implemented in the
development of a complementary application for the clinical diagnosis or monitoring
of a patient under the supervision of a specialist.
* Tecnológico Nacional de México, Campus Apizaco, Departamento de Sistemas Computacionales, México,
podriaservirte@[Link], [Link]@[Link], [Link]@[Link]
** Instituto Nacional de Astrofísica, Óptica y Electrónica, Departamento de Ciencias y Tecnologías Biomédicas,
México, kargaxxi@[Link]
*** Universidad Autónoma de Asunción, Facultad de Ciencias de la Salud, Departamento de Neuropsicología,
Paraguay, ttalavera@[Link]
14
1. INTRODUCTION
Dysphonia is a voice disorder characterized by the abnormal loss of typical voice quality
due to functional or organic disturbances in the larynx. It can be classified into three primary
categories: Functional, Organic, and Mixed (Behlau & Pontes, 1989).
Functional dysphonia, often associated with emotional stress, poor vocal habits, or
excessive voice use in daily activities, is typically accompanied by symptoms such as
intermittent hoarseness, vocal projection difficulties, and occasional throat discomfort or
tension due to vocal fatigue.
Conversely, organic dysphonia results from physical laryngeal disorders, including
injuries, infections, tumors, or other medical conditions that directly impact the laryngeal
structure. It tends to present as persistent hoarseness, significant alterations in vocal quality,
and occasional throat pain or discomfort.
Mixed dysphonias represent a combination of both organic and functional factors,
resulting in diverse symptoms, including variable hoarseness, shifts in vocal quality,
discomfort, occasional vocal fatigue, and more. Recognizing these distinctions is imperative
in clinical practice, as they necessitate distinct therapeutic approaches.
Another indicator frequently utilized to differentiate vocal disorders pertains to the
biomechanics of phonation, which distinguishes between hyperfunction and hypofunction.
Hyperfunctional dysphonia arises from excessive engagement of the muscles involved in
glottic closure, affecting the ventricular bands, resulting in a tense, rough, high-frequency
voice with increased intensity (López, 1997). Conversely, hypofunctional dysphonia is
characterized by laryngeal muscle weakness, incomplete glottis closure, and a breathy, low-
intensity voice (López, 2000). The former is typically associated with functional dysphonias,
while the latter is linked to organic dysphonias (Farias, 2016). The recent COVID-19
pandemic has led to both hyperfunctional and hypofunctional voice alterations.
The causes of dysphonia are multifaceted and encompass factors such as viral invasion
of the glottic epithelium, as observed in infections, which can result in damage. Research by
Hoffmann et al. (2020) suggests that the ACE2 (Angiotensin-Converting Enzyme 2) protein,
used as a receptor by certain viruses like SARS-CoV-2, plays a pivotal role in these vocal
fold alterations. Moreover, Descamps et al. (2020) has demonstrated the expression of ACE2
in the epithelial cells of the vocal folds.
During the COVID-19 pandemic, dysphonia has been reported as a prevalent symptom,
especially in patients with laryngeal inflammation, even in post-infection stages.
Additionally, it can also arise in patients who have undergone tracheal intubation (Verdaguer
et al., 2008), where structural changes in the larynx can impact vocal function. Prolonged
intubation can lead to mild to severe damage.
Dysphonia requires therapy from an otolaryngologist or a speech therapist, who need
objective tools to provide solutions (Behlau et al., 2005), which can be developed in the field
of artificial intelligence, where machine learning algorithms are implemented to classify data
for a specific purpose, and are frequently used in medicine to develop diagnostic or
rehabilitation systems.
In clinical practice, it is essential to distinguish between hypofunctional and
hyperfunctional dysphonia, as they require different rehabilitation schemes. Additionally, a
patient may switch from one to another due to compensatory phonation mechanisms.
15
In this paper, the harmonic-to-noise ratio (HNR), cepstral peak prominence smoothed
(CPPS), zero-crossing rate (ZCR) and the means of the Mel frequency cepstral coefficients
(2-19) are used to perform a multiclass classification of euphonic, hyperfunctional and
hypofunctional voices using six machine learning algorithms: Random Forest, K-Nearest
Neighbors, Logistic Regression, Decision Trees, Support Vector Machines and Naive Bayes.
The performance of these algorithms in identifying the three voice types through the
mentioned acoustic parameters was evaluated using bootstrap.632. The results obtained from
this research can be useful in developing a complementary application for clinical diagnosis
or patient monitoring under specialist supervision.
The second section discusses some related works to observe how the problem has usually
been addressed. The third section describes the methods and materials used in developing
this project. The fourth section presents the implementation. The fifth section analyzes the
results obtained. Finally, the sixth section presents the conclusions and future work.
2. RELATED WORKS
Several projects have been developed aiming to identify disorders that affect voice
quality, even working with the same databases used in the present project. However, they
focus on binary classification between healthy and pathological voices, that is, identifying
whether a person is sick or not.
Verde et al. (2019) estimated F0, Jitter, Shimmer, and HNR from 2003 samples (796
healthy voices and 1207 pathological voices) as input to the classifier. The Boosted Trees
algorithm is employed, using an 80-20 proportion for the training and testing phases,
obtaining metrics of 83% and 86% for sensitivity and specificity, respectively. The authors
clarify that the classification is binary. They use three databases: Massachusetts Eye and Ear
Infirmary (MEEI), Saarbruecken Voice Database (SVD), and VOice ICar fEDerico II
(VOICED).
Chen & Chen (2022) used a deep neural network for classification, extracting MFCC
features (12 coefficients) from 114 samples of the VOICED database and obtaining metrics
of sensitivity, specificity, precision, accuracy, and F1 with 97.8%, 99.4%, 99.4%, 98.6%,
and 98.4%, respectively. It should be noted that the classification is binary.
Altayeb & Al-Ghraibah (2022) explored different voice feature extraction methods,
including Mel frequency cepstral coefficients, zero-crossing rate, and discrete wavelet
transform. The support vector machine algorithm was employed for classification, achieving
100% accuracy using a group of MFCC features (12 coefficients) and kurtosis, based on the
VOICED database. The classification is also binary, as in previous projects, but the
difference lies in conducting separate classifications for each disorder with a healthy voice,
rather than distinguishing between pathological voices.
Hassan et al. (2020) used a dataset corresponding to 80 people (60 healthy and 20 with
pathology), with audio samples of coughs, breathing, and voice. An LSTM-type RNN was
used on the features: Spectral centroid, spectral roll-off, zero-crossing rate, and MFCC. A
70/30 ratio was adopted between training and testing. The area under the curve (AUC)
reached around 97.4%, 98.8%, and 84.4% when classifying cough, breathing, and voice,
respectively.
16
Radha et al. (2021) used the parameters chroma STFT, spectral centroid, spectral
bandwidth, MFCC, roll-off, and zero-crossing rate for Parkinson's disease detection, as
speech deterioration is an early indicator in this disease. A classification method based on
convolutional neural networks, artificial neural networks, and hidden Markov model is used
to distinguish samples of Parkinson's patients and healthy individuals. The artificial neural
network-based model achieved a recognition rate of 96%.
3. METHODS AND MATERIALS
3.1. Databases
The dataset used throughout this project consists of samples from two databases:
Saarbruecken Voice Data Base (Woldert-Jokisz, 2007) and Voice ICar fEDerico II (Cesari et
al., 2018). Exactly 331 samples were taken, 208 from SVD and 123 from VOICED,
corresponding to a total of 193 individuals, of whom 55 have hypofunction, 64 have
hyperfunction, and 74 have euphony. All audios are sustained phonations of the vowel /a/. The
samples were preprocessed since SVD samples have a sampling frequency of 50 kHz and a
duration of 1 second, while VOICED samples have a sampling frequency of 8 kHz and a duration
of approximately 4 seconds. To homogenize the samples, those with a longer duration were
divided, setting all to 1 second, and the sampling frequency was modified to 8 kHz, equalizing
the frequency range present. The number of samples for each class was standardized to 190,
which is the maximum number of samples that could be obtained for the hypofunction class.
Therefore, the dataset used is composed of 570 samples in total, 190 samples with euphony, 190
with hyperfunction, and 190 with hypofunction.
3.2. Acoustic Parameters
3.2.1. Means of the Mel Frequency Cepstral Coefficients
Mel Frequency Cepstral Coefficients (MFCC) are a representation of the audio signal that is
commonly used in speech processing and voice recognition. To obtain these coefficients, a Short-
Time Fourier Transform (STFT) is applied to a voice signal. Then, the resulting power is mapped
onto a logarithmic Mel-frequency scale, and the cepstral coefficients are calculated from the
resulting spectrum. Since the MFCC are calculated on a logarithmic scale, it fits better with
human auditory perception, which is more sensitive to differences in frequency in the low-
frequency range than in the high-frequency range (Rivera et al., 2022). In general, MFCC are
used as an efficient representation of a voice signal, which may be relevant for identifying voices
with euphonia, hyperfunction, and hypofunction. It is worth noting that only the means of the
coefficients are used, and not the complete vectors of the coefficients, as using the means reduces
the dimensionality of the data and eliminates redundancy. This is due to the fact that the samples
belong to sustained phonations of the vowel /a/ (Flórez-Gómez et al., 2022), so it would be
expected that a voice with euphonia (which is characterized by being balanced and harmonic)
possibly presents more constant or similar values throughout the vectors of each coefficient,
while voices with hypofunction and hyperfunction possibly present more outliers and variability
that would raise the values of the means, which is considered relevant for classification.
17
3.2.2. Cepstral Peak Prominence-Smoothed (CPPS)
Cepstral Peak Prominence-Smoothed represents the difference between the most prominent
cepstral peak, which corresponds to the first rahmonic, and the point with the same quefrency
on the regression line through the smoothed cepstrum. Studies indicate that this parameter
reflects the degree of dysphonia (Núñez-Batalla et al., 2019). Currently, in clinical practice,
CPPS is the most widely used acoustic parameter for measuring vocal quality levels.
3.2.3. Zero Crossing Rate (ZCR)
Zero Crossing Rate (ZCR) is a value between 0 and 1 that indicates an average number of
times the signal crosses the X-axis in defined intervals (Celdrán, 2015). Zero Crossing Rate can
be useful in identifying differences between voices with euphonia and those with hyperfunction
and hypofunction, as they may have a different zero crossing rate due to changes in vocal fold
vibration.
3.2.4. Harmonic-to-Noise Ratio (HNR)
Harmonic-to-Noise Ratio is an evaluation of the ratio between the periodic and non-periodic
components comprising a segment of vocal speech. The periodic component comes from vocal
fold vibration, while the non-periodic component comes from glottal noise, expressed in dB. The
evaluation of the ratio between these components reflects speech efficiency; that is, the greater
the flow of air expelled from the lungs in vibrational energy of the vocal folds, the higher the
HNR. A low HNR denotes an asthenic voice and dysphonia.
3.3. Machine learning models
The six machine learning models that have been employed to address the task of multiclass
classification of the three voice classes in question are described below.
3.3.1. Random Forest
Random Forest is an ensemble model that uses multiple decision trees to perform
classification. Each tree is trained with a random subsample of the data and a random selection
of features. To classify an input object x, its position is evaluated in each tree and a class label is
assigned based on the results from all the trees. The final classification is determined by voting
the results of the individual trees (Schonlau & Zou, 2020).
3.3.2. K Nearest Neighbors
It is a classification model based on distance. The idea is that an object is classified based on
the labels of its k nearest neighbors in the feature space. The value of k is chosen beforehand and
can vary depending on the problem.
For an input object x, the distance between x and each of the training points is calculated, and the
k closest points are selected. The class of the input object is determined by majority voting among
the k classes of the nearest neighbors (Taunk et al., 2019).
18
3.3.3. Logistic Regression
Logistic regression is a linear classification model that uses the logistic function to predict the
probability of belonging to a class. It can be extended for multiclass classification using
techniques such as One-vs-Rest or SoftMax. In the One-vs-Rest approach, a logistic regression
model is trained for each class, while in the SoftMax approach, a SoftMax activation function is
used to assign probabilities to each class. Multiclass logistic regression uses a loss function such
as Categorical Cross-Entropy to train the model and adjust the parameters (Daniels & Minot,
2019).
3.3.4. Decision Trees
It is a classification model based on creating a tree of decision rules. Each node of the tree
represents a decision rule based on a feature, and each branch represents a possible outcome of
that rule. The tree is constructed from the training data, and the final classification is performed
by traversing the tree from top to bottom until reaching a leaf that contains the class label
(Charbuty & Abdulazeez, 2021).
3.3.5. Support Vector Machines
Support vector machines are classification models that attempt to find the best separation
between classes by creating a hyperplane that maximizes the distance between the closest
samples of each class. A kernel function is used to project the data into a higher-dimensional
feature space to find the optimal hyperplane (James et al., 2013).
3.3.6. Naive Bayes
Naive Bayes is a probabilistic classification model based on Bayes' theorem and the naive
assumption that all features are independent of each other. Using training data, the model
calculates the probability of an object belonging to each class using Bayes' theorem (Murphy,
2006).
3.4. Model Evaluation Using Bootstrap.632
Bootstrap.632 is a technique for evaluating machine learning models that uses bootstrap
sampling to generate multiple random samples with replacement from the training dataset and
trains a model on each of them. The model is then evaluated on a test dataset that was not used
to train the model, and the evaluation of the models is adjusted using a correction based on the
probability of a particular data point being in the training or test set in each bootstrap sample.
This technique produces a more accurate estimate of the model's performance on new and
unknown data (Efron, 1983).
4. IMPLEMENTATION
The data analysis and machine learning models training were programmed in Python,
using various libraries such as NumPy, Pandas, and Scikit-learn. Additionally, the
19
Parselmouth library was used for the analysis of acoustic signals and the extraction of
acoustic parameters. All the code was executed on Google Colab.
In order to obtain better results, hyperparameter optimizations were performed on the
machine learning models, using the grid search method, which resulted in the following
configurations:
1. For the Random Forest model, the entropy criterion was chosen to evaluate the
quality of the partitions at each node. The maximum depth of the trees in the random
forest was set to 30 levels. All available features were considered when searching for
the best partition at each node. A minimum of 2 samples were required to split an
internal node. The random forest consisted of a total of 500 decision trees,
2. For the K Nearest Neighbors model, the ball tree algorithm was used to calculate the
distances between instances. Three nearest neighbors were considered for
classification, and the Manhattan distance (L1) was used as the distance metric.
During classification, the nearest neighbors have more influence, as they are
weighted inversely proportional to the distance,
3. For the Decision Trees model, a split quality measure based on the 'log_loss' criterion
was used. The tree was configured with a maximum depth of 20 levels to prevent
overfitting. All features were considered, with the number of maximum features set
to the logarithm with a base of 2 of the total number of features. Moreover, a
minimum of 2 samples were required in a node to perform an additional split. The
'best' strategy was employed to select the optimal split based on the defined quality
criterion.
4. For the Support Vector Machines model, a combination of hyperparameters with
C=10, gamma=0.2, and kernel='rbf' achieved the best results in terms of accuracy.
This indicates that a high value of C was used, prioritizing precise classification even
with a narrower separation margin. Additionally, a moderate value of gamma (0.2)
was chosen to consider the influence of nearby training points on the decision
boundary. The 'rbf' kernel was used to capture nonlinear relationships between the
data,
5. For the Logistic Regression model, the optimal combination of hyperparameters was
found to be C=10, penalty='l2' (Ridge regularization), and solver='newton-cg'. This
indicates that a relatively weak regularization was chosen to prevent overfitting by
adding the square of the coefficients. The 'newton-cg' solver was used, which is
suitable for multiclass optimization problems and relies on the conjugate Newton
method to optimize the model's objective function,
6. For the Naive Bayes model, it was found that the optimal configuration includes a
smoothing term for the covariance matrix of the Gaussian distributions equal to
1𝑒 −09.
5. RESULTS ANALYSIS
This section presents the results of the machine learning models, which have undergone
hyperparameter optimization and bootstrap.632 evaluation (see Table 1). Confidence
intervals for 500 iterations based on the accuracy metric of each model were obtained for
the multiclass classification task of voices with euphonia, hypofunction, and hyperfunction.
20
Tab. 1. Confidence intervals of each machine learning model
Model Confidence interval
Random Forest 86-90%
K Nearest Neighbors 87-92%
Support Vector Machines 84-91%
Decision Trees 79-85%
Logistic Regression 58-65%
Naïve Bayes 53-61%
The distribution histograms obtained after applying the Bootstrap.632 method for the
three best performing models is shown below.
5.1. Random Forest Results
The confidence interval obtained suggests that the model has a high probability of obtaining
an accuracy between 86% and 90% on new data sets, which is an indicator that the model may
be useful for the task of classifying voices with euphony, hypofunction and hyperfunction. Figure
1 shows that in more than 120 of the 500 iterations performed there is a concentration in
accuracies of 88%. Table 2 shows the densities of the Random Forest model accuracies.
Fig. 1. Histogram of predictions and confidence interval of the Random Forest model
Tab. 2. Densities of the Random Forest model accuracies
Accuracy Density
84.35% 2
85.06% 7
85.77% 18
86.48% 53
87.19% 89
87.90% 131
88.61% 97
89.32% 64
90.03% 31
90.74% 8
21
5.2. K-Nearest Neighbors Results
The 87% to 92% confidence interval indicates that if multiple random samples are taken from
the data set, the accuracy of the model is expected to be within this range in most cases.
Furthermore, the fact that the range is narrow indicates that the model is robust and that its
performance is unlikely to vary significantly across different data sets. Figure 2 shows that much
of the accuracies fall between 89% and 91% of the 500 iterations performed, with its highest
concentration at 90% in more than 130 of the total iterations. Table 3 shows the densities of the
K nearest neighbors model accuracies.
The results obtained suggest that the K nearest neighbors model is a good candidate for voice
classification in the three classes mentioned above, and that its accuracy is high enough to be
considered useful in practical applications.
Fig. 2. Histogram of predictions and confidence interval of the K-Nearest Neighbors model
Tab. 3. Densities of the K-Nearest Neighbors model accuracies
Accuracy Density
85.25% 2
86.10% 3
86.96% 9
87.81% 42
88.67% 91
89.52% 138
90.38% 134
91.23% 57
92.09% 18
92.94% 6
5.3. Support Vector Machine Results
The confidence interval obtained for this model, ranging from 84% to 91%, may indicate
that the model's performance is consistent and stable. It suggests that the model is able to
accurately classify most of the samples. Figure 3 shows that there are a large number of
22
accuracies that fall around 87%; however, many points fall within a higher percentage range,
going from 88% to 90%. The model is able to classify all three speech classes with
acceptable accuracy. Table 4 shows the densities of the Support Vector Machine model
accuracies.
Fig. 3. Histogram of predictions and confidence interval of the Support Vector Machine model
Tab. 4. Densities of the Support Vector Machine model accuracies
Accuracy Density
82.99% 9
83.99% 20
84.99% 53
85.99% 62
86.99% 118
87.99% 97
88.99% 88
89.99% 29
90.99% 17
91.99% 7
6. CONCLUSIONS AND FUTURE WORK
According to the presented results, it can be concluded that the K Nearest Neighbors
model performs best for the multiclass classification task of voices with euphony,
hypofunction, and hyperfunction, and could be considered for the development of
complementary tools that may be useful in the objective evaluation of patients under the
supervision of a specialist, by identifying the three voice classes and assigning an appropriate
rehabilitation scheme. The KNN model presents a narrower confidence interval and a higher
accuracy (87-92%) than other models such as Random Forest (86-90%) and Support Vector
Machines (84-91%). Decision trees (79-85%) are above the 70% threshold that is considered
acceptable for many classification problems; however, it is not the best option. On the other
hand, logistic regression, and Naive Bayes showed less satisfactory performance with
23
confidence intervals that do not exceed 65%, indicating that these models may not be the
most suitable for the task at hand. It should be noted that the classifications performed are
multiclass, which is relevant considering that hypofunction and hyperfunction are two
common types of dysphonia, and although they have individual characteristics, they also
share some features.
It is important to note that these results are valid for the hyperparameter settings and the
dataset used in this study, and further experiments should be conducted to validate these
results for different configurations and datasets.
Data availability
The set of voice samples with euphonia, hypofunction, and hyperfunction used in this
study, which belong to the VOICED database, are publicly available under an Open Data
Commons Attribution License on PhysioNet's website at
[Link] Additionally, voice samples from the
Saarbruecken Voice Database are freely accessible to the public, provided by the Institute
of Phonetics at the University of Saarland, and can be found at
[Link]
Acknowledgments
The authors thank the reviewers for their helpful suggestions that helped improve the
quality of this article.
REFERENCES
Altayeb, M., & Al-Ghraibah, A. (2022). Classification of three pathological voices based on specific features
groups using support vector machine. International Journal of Electrical and Computer Engineering
(IJECE), 12(1), 946-956. [Link]
Behlau, M., & Pontes, P. (1989). Avaliação Global da Voz. Editora Paulista Publicações Médicas.
Behlau, M., Madazio, G., Feijó, D., Azevedo, R., Gielow, I., & Rehder, M. (2005). Perfeccionamiento vocal y
tratamiento fonoaudiológico de las disfonías. In M. Behlau (Eds.), Voz: O livro do especialista. Thieme
Revinter.
Celdrán, E. M. (2015). Naturaleza fonética de la consonante ‘ye’en español. Normas: revista de estudios
lingüísticos hispánicos, 5, 117-131. [Link]
Cesari, U., De Pietro, G., Marciano, E., Niri, C., Sannino, G., & Verde, L. (2018). A new database of healthy
and pathological voices. Computers & Electrical Engineering, 68, 310-321.
[Link]
Charbuty, B., & Abdulazeez, A. (2021). Classification based on decision tree algorithm for machine learning.
Journal of Applied Science and Technology Trends, 2(01), 20-28. [Link]
Chen, L., & Chen, J. (2022). Deep neural network for automatic classification of pathological voice signals.
Journal of Voice, 36(2), 288.e15-288.e24. [Link]
Daniels, L., & Minot, N. (2019). An introduction to statistics and data analysis using Stata®: From research
design to final report. Sage Publications.
Descamps, G., Verset, L., Trelcat, A., Hopkins, C., Lechien, J. R., Journe, F., & Saussez, S. (2020). ACE2 protein
landscape in the head and neck region: the conundrum of SARS-CoV-2 infection. Biology, 9(8), 235.
[Link]
Efron, B. (1983). Estimating the error rate of a prediction rule: improvement on cross-validation. Journal of the
American Statistical Association, 78(382), 316-331. [Link]
Farias, P. (2016). Guía clínica para el especialista en laringe y voz. Librería Akadia Editorial.
24
Flórez-Gómez, A. F., Orozco-Arroyave, J. R., & Roldán-Vasco, S. (2022). Correlación entre espacios de
características acústicas del habla y trastornos clínicos de la voz en pacientes con disfagia. TecnoLógicas,
25(53), e2220. [Link]
Hassan, A., Shahin, I., & Alsabek, M. B. (2020). COVID-19 detection system using recurrent neural networks.
2020 International conference on communications, computing, cybersecurity, and informatics (CCCI)
(pp. 1-5). IEEE. [Link]
Hoffmann, M., Kleine-Weber, H., Schroeder, S., Krüger, N., Herrler, T., Erichsen, S., Schiergens, T. S., Herrler,
G., Wu, N.-H., Nitsche, A., Müller, M. A., Drosten, C., & Pöhlmann, S. (2020). SARS-CoV-2 cell entry
depends on ACE2 and TMPRSS2 and is blocked by a clinically proven protease inhibitor. Cell, 181(2),
271-280.e8. [Link]
James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning. Springer.
López, J. A. P. (1997). Los trastornos de la voz en el personal docente de logroño. Estudio de la voz en los
profesionales de la enseñanza. (Doctoral dissertation, Universidad de Navarra).
López, J. A. P. (2000). Estudio de la prevalencia de los trastornos de la voz en el personal docente de Logroño.
Zubía, 12, 111-145.
Murphy, K. P. (2006). Naive bayes classifiers. University of British Columbia, 18(60), 1-8.
Núñez-Batalla, F., Cartón-Corona, N., Vasile, G., García-Cabo, P., Fernández-Vanes, L., & Llorente-Pendás, J.
L. (2019). Validez de las medidas del pico cepstral para la valoración objetiva de la disfonía en sujetos
de habla hispana. Acta Otorrinolaringológica Española, 70(4), 222-228.
[Link]
Radha, N., Sachin Madhavan, R. M., & Sameera holy, S. (2021). Parkinson’s Disease detection using Machine
Learning Techniques. International Journal of Early Childhood Special Education (INT-JECSE), 30(2),
543. [Link]
Rivera, M. A. B., Flores, P. M. Q., Loaiza, R. E. P., & Rivera, L. G. (2022). Analysis of audio signals using deep
learning algorithms applied to COVID diagnostic systems. 2022 IEEE Mexican International Conference
on Computer Science (ENC) (pp. 1-6). IEEE. [Link]
Schonlau, M., & Zou, R. Y. (2020). The random forest algorithm for statistical learning. The Stata Journal, 20(1),
3-29. [Link]
Taunk, K., De, S., Verma, S., & Swetapadma, A. (2019). A brief review of nearest neighbor algorithm for
learning and classification. 2019 International Conference on Intelligent Computing and Control Systems
(ICCS) (pp. 1255-1260). IEEE. [Link]
Verdaguer, J. M., Górriz, C., Prim, M. P., del Palacio, A. J., Gavilán, J., & de Diego, J. I. (2008). Análisis de los
cambios en el espectrograma tras la intubación endotraqueal. Acta Otorrinolaringológica Española,
59(5), 217-222. [Link]
Verde, L., De Pietro, G., Alrashoud, M., Ghoneim, A., Al-Mutib, K. N., & Sannino, G. (2019). Leveraging
artificial intelligence to improve voice disorder identification through the use of a reliable mobile app.
IEEE Access, 7, 124048-124054. [Link]
Woldert-Jokisz, B. (2007). Saarbruecken voice database. Computer Science.
25