A novel convolutional neural network based dysphonic voice detection algorithm using chromagram

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 5, October 2022, pp. 5511~5518
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i5.pp5511-5518  5511
Journal homepage: http://ijece.iaescore.com
A novel convolutional neural network based dysphonic voice
detection algorithm using chromagram
Rumana Islam1
, Mohammed Tarique2
1
Department of Electrical and Computer Engineering, Faculty of Engineering, University of Windsor, Windsor, Canada
2
Department of Electrical and Computer Engineering, College of Engineering and Technology, University of Science and Technology of
Fujairah, Fujairah, United Arab Emirates
Article Info ABSTRACT
Article history:
Received Aug 1, 2021
Revised May 24, 2022
Accepted Jun 20, 2022
This paper presents a convolutional neural network (CNN) based non-
invasive pathological voice detection algorithm using signal processing
approach. The proposed algorithm extracts an acoustic feature, called
chromagram, from voice samples and applies this feature to the input of a
CNN for classification. The main advantage of chromagram is that it can
mimic the way humans perceive pitch in sounds and hence can be
considered useful to detect dysphonic voices, as the pitch in the generated
sounds varies depending on the pathological conditions. The simulation
results show that classification accuracy of 85% can be achieved with the
chromagram. A comparison of the performances for the proposed algorithm
with those of other related works is also presented.
Keywords:
Chromagram
Classification
Deep learning
Dysphonia
Machine learning
Voice pathology
This is an open access article under the CC BY-SA license.
Corresponding Author:
Rumana Islam
Department of Electrical and Computer Engineering, Faculty of Engineering, University of Windsor
401 Sunset Avenue, ON N9B 3P4, Canada
Email: islamq@uwindsor.ca
1. INTRODUCTION
Voice disability is a barrier to effective human speech communication. Primarily, voice disability
occurs due to improper function of the components that constitute the human voice generation system
[1]–[3]. Various voice disabilities have been reported in the literature. The American speech-language-
hearing association (ASHA) has identified dysphonia [4] as the most common voice pathology. Dysphonia
refers to abnormal voices [5], [6] that can develop suddenly (or gradually) over time. Commonly, dysphonia
is caused by inflamed vocal folds that cannot vibrate properly to produce normal voice sounds [7], [8].
Both invasive and non-invasive methods are used for pathological voice detection. In invasive
methods, physicians insert probe into the patient’s mouth using endoscopic procedures, namely,
laryngoscopy [9], stroboscopy [10], and laryngeal electromyography [11]. In non-invasive methods, audio
signals are used for voice pathology detection. Audio signals including cough, breathing, and voice have
been popularly used in many applications including telecommunication, robot control [12], domestic
appliance control, data entry, voice recognition, disease diagnosis [13], [14], and speech-to-text conversion
[15], [16]. Voice signals are mainly used to implement non-invasive voice pathology detection algorithms
[17]. The two-fold objectives of these methods are i) to reduce the discomfort of a patient and ii) to assist the
clinicians in preliminary diagnosis of voice pathology. In this method, voice samples are collected in a
controlled environment and acoustic features are extracted from the voice samples. The next step is to
classify the samples into two categories namely normal (i.e., healthy) and pathological using a classifier
algorithm. Numerous classifier algorithms have been suggested in the literature for pathological voice
detection. Recently, machine learning and deep learning algorithms have drawn considerable attention from

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5511-5518
5512
researchers [18]–[22]. Specifically, deep learning algorithms demonstrate promising results and provide high
accuracy in voice pathology detection [19]–[22].
Various voice disability detection algorithms have been published in the literature. In [23], the
authors have used a spectrogram to detect voice pathology. They have suggested minimizing the effects of
jitter, shimmer, and harmonic-to-noise ratio (HNR) to improve the classification accuracy with the
spectrogram. In [24], eight-voice pathologies have been investigated, and the results show that deep neural
network (DNN) based classifier achieves a high accuracy (94.26% and 90.52% in male and female subjects,
respectively).
Vocal disorders, namely neoplasm, phono-trauma, and vocal palsy have been investigated in [25].
The authors used a dense net recurrent neural network (DNRNN) in their work. The results show that the
DNRNN algorithm achieves an accuracy of 71%. In another similar work [26], multiple neural networks
have been used to detect voice pathology. The authors have used multilayer perceptron neural network
(MLPNN), general regression neural network (GRNN), and probabilistic neural network (PNN) in this work,
and they achieved the highest accuracy with the MLPNN. Some researchers suggested using multiple voice
features to increase accuracy. For example, the researchers in [27] have used six-voice features, namely jitter,
shimmer, harmonic-to-noise ratio (HNR), soft phonation index (SPI), amplitude perturbation quotient (APQ),
and relative average perturbation (RAP). They have achieved the classification accuracy of 95.20% and
84.20% with generalized method of moments (GMM) and artificial neural network (ANN), respectively.
Support vector machine (SVM) and radial basis function neural network (RBFNN) have been used
in [28] to detect voice pathology. In their work, the authors have also used several features, and they
achieved an accuracy of 91% with RBFNN. On the other hand, SVM achieved an accuracy of 83%.
Stuttering voice has been addressed in [29]. The authors developed a classifier that can detect stuttered
speech in their work. The results presented in their work show that these algorithms can detect stuttered
voices with an accuracy of 85% and 78% for males and females, respectively. Four voice attributes, namely
roughness, breathiness, asthma, and strain have been considered in [30]. The proposed algorithm used these
features for the classification by using a feed-forward neural network (FFNN) and the algorithm achieved an
average F-measure of 87.25%.
Some researchers claim that spectrogram is the most suitable voice feature to detect voice pathology
because it traces different frequencies and their occurrences in time. For example, the authors have used
spectrogram to detect pathological voice disorder due to vocal cord paralysis in [31]. The spectrograms of
pathological and normal speech samples are applied to the input of a convolutional deep belief network
(CDBN). The authors have achieved 77% and 71% accuracy for CNN and CDBN, respectively.
Dysphonic voice detection using a pre-trained CNN has been presented in [32]. The results show
that the proposed method can detect dysphonic voices with an accuracy of 95.41%. To detect dysphonic
voice, a new marker called the dysphonic marker index (DMI) has been introduced in [33]. The marker
consists of four acoustic parameters. The authors have employed a regression algorithm to relate this marker
to discriminate pathological voices from healthy ones. A novel computer-aided pathological voice
classification system is proposed in [34]. In this work, the authors have used a deep-connected ResNet for
classification.
There are two significant limitations of the above-mentioned related works. One limitation is that
none of the works considers the way human auditory system perceives the pitch contained in the sounds.
Another limitation is that these works use classifiers that overwhelm the system with a substantial
computational burden. The limitations mentioned above are overcome in this work by using the chromagram
feature and a CNN. Chromagram has been commonly used to detect musical sounds, as far as our knowledge
goes, we are the first to use the chromagram in pathological voice detection. The main contributions of this
work are: i) developing a novel pathological voice detection algorithm based on signal processing and a deep
learning approach, ii) extracting the chromagram feature from the voice samples and converting them into a
suitable form for classification, iii) achieving a high classification accuracy without overwhelming the system
with a vast computation burden, iv) providing a detailed performance analysis of the proposed system in
terms of classification matrix, precision, sensitivity, specificity, and F1-score, and v) comparing the
performances of the proposed algorithm with other related works to demonstrate its effectiveness. The rest of
the paper is organized as follows: materials and methods are presented in section 2. The results are presented in
section 3, and the paper is concluded with section 4.
2. MATERIALS AND METHODS
This investigation uses normal and dysphonic voice samples that are collected from the Saarbrucken
voice database (SVD) [35]. The SVD database contains 687 normal (i.e., healthy) samples and 1,356
pathological samples. The samples contain the recordings of vowels and sentences. In this investigation, the

Int J Elec & Comp Eng ISSN: 2088-8708 
A novel convolutional neural network based dysphonic voice detection algorithm … (Rumana Islam)
5513
sustained phonation of the vowel ‘/a/’ is used. The main reason is that a speaker can maintain a steady
frequency and amplitude at a comfortable level during the voice generation of the vowel ‘/a/’. Moreover, it is
free of articulatory and other linguistic confounds that often exist with other common speech components,
including sentences and running speeches. Figure 1 shows the plots for a normal voice sample and a
dysphonic voice sample that are randomly selected from the SVD database. It is depicted in the figure that
the dysphonic voice samples demonstrate irregular distortion both in terms of magnitude and shape compared
to that of the healthy sample.
Figure 1. The normal and dysphonic voice samples
In this work, the chromagram audio feature has been used for the classification purpose. The
chromagram is derived based on the principle of human’s auditory perception of pitch in the sound. The pitch
is represented as shown in Figure 2. Generally, humans perceive the pitch as extending along a scale from
low to high on a circular scale traversing in the clockwise direction as shown in Figure 2(a). In addition, the
pitch has a linear scale too. To accommodate both linear and circular dimensions, Shepard introduced the
concept of chroma in [36]. He suggested that the human auditory system’s pitch perception is better
represented as a helix as shown in Figure 2(b). According to Shepard, the pitch, 𝑝 perceived by a human, can
be expressed in terms of chroma, 𝑐, and tone-height, ℎ by (1).
𝑝 = 2ℎ+𝑐
(1)
Patterson generalized Shepard’s model in [37] to find the pitch by using (2).
𝑓 = 2ℎ+𝑐
(2)
The chroma spectrum, 𝑆(𝑐) is defined as the measure of the strength of a signal with a given value of
Chroma. This is analogous to the standard Fourier power spectrum of a signal. Then, the chroma spectrum is
extended in the time domain to create a time-frequency distribution (TFD) 𝑆(𝑐, 𝑡). This distribution is called
the chromagram and it represents a joint distribution of signal strength over the variables of chroma and time.
The chromagram is generated by two major steps in this work, namely frame segmentation and
feature calculation. Rather than using a uniform frame size, a beat-synchronous frame segmentation is used
in this work. This allows the frame sampling to track the rhythm of the voice. The second step of the
algorithm is the feature calculation. A 12-element representation of the spectral energy has been used to
calculate the chroma vector. Then, the chroma vector is computed by grouping the discrete Fourier transform
(DFT) coefficients of a short-term window into 12 bins. Each bin represents 12 equal-tempered pitch classes
of semitone spacing. Each bin produces the mean of log-magnitudes of the respective DFT coefficients
defined by (3).
𝑣𝑘 = ∑
𝑋𝑖(𝑛)
𝑁𝑘
𝑛∈𝑆𝑘
, kϵ0 …11 (3)

 ISSN: 2088-8708
5514
where, 𝑆𝑘 is a subset of the frequencies that correspond to the DFT coefficients and 𝑁𝑘 is the cardinality
of 𝑆𝑘. This computation results in a matrix, 𝑉 with elements 𝑉𝑘,𝑖, where indices 𝑘 and 𝑖 represent pitch-class
and frame-number, respectively. The matrix, 𝑉𝑘,𝑖 is represented in a suitable form to produce the
chromagram. The detailed algorithm of computing the chroma vector is shown in Algorithm 1. The
chromagram of normal and dysphonic voices is plotted in Figure 3. This figure shows the distinct differences
between the chromagram of the normal and dysphonic voices. For example, the chromagram of the normal
voice shows a few dominant coefficients, which are stable for a short period. However, the chromagram of
dysphonic voice is noisier compared to that of normal voice.
(a) (b)
Figure 2. The pitch representation (a) circular configuration of pitch and (b) Shepard’s helix model [37]
Algorithm 1. Generating the chromagram of voice signals
/*Set the initial value of the parameters */
Set the number of tone heights 𝒉 = 𝟏𝟐;
Set the number of bin 𝑩 = 𝟏𝟐;
Set the fundamental frequency 𝐟𝟎 = 𝟓𝟓 Hz;
Load the sound data into vector 𝐗;
Read the sampling rate from the soud file and store into variable 𝑭𝒔;
Normalize the data by 𝑿𝒏𝒐𝒓𝒎 =
𝑿
𝐦𝐚𝐱 (𝑿)
;
Set window size 𝑾;
Calculate the number of frame 𝑵 =
𝒍𝒆𝒏𝒈𝒕𝒉(𝑿)
𝑾
;
/* Set the frequency range to compute the spectrogram*/
Set minimum frequency 𝒇𝒎𝒊𝒏 = 𝟎;
Calculate the maximum frequency 𝒇𝒎𝒂𝒙 =
𝑭𝒔
𝑾
;
Calculate the chromatic scale 𝒇𝒊 = 𝒇𝟎𝟐𝒊/𝒉
Compute the discrete-time Fourier transform of the signal vector 𝑿(𝒌)
/*Determine the log-magnitudes of the respective DFT co-efficients */
while 𝒊 ≤ 𝑵
while 𝒌 ≤ 𝑩
𝒗𝒌 = ∑𝒏𝝐𝑺𝒌
𝑿𝒊(𝒏)
𝑵𝒌
end while
Convert 𝒗𝒌 into 𝑽𝒌,𝒊
end while
Generate chromagram 𝑪(𝒎, 𝒕)
In this work, a CNN is employed as the binary classifier. The CNN model presented in [38] is used
as the base to implement the classifier. The CNN model includes two networks namely feature extraction
network and classifier network. The input data (i.e., chromagram) is applied to the feature extraction
network. The extracted feature map is applied to the classification neural network. The feature extractor
network consists of piles of convolutional layer and pooling layer pairs as shown in Figure 4. To avoid
computational burden on the system, one convolutional layer is used as the feature extractor network in this
work. The CNN model uses 20 convolutional filters of size 9 × 9. The feature map produced by the
convolutional filters is processed by an activation function. The rectified linear unit (ReLU) function is used
for this purpose. The output produced by the convolutional layer is then passed through the pooling layer. In
this work, a 2 × 2 matrix for pooling is used to find the mean value from the input data. The hidden layer has
100 nodes that also use the ReLU activation function. The output layer of the CNN contains a single node as
the decision made by the classifier is binary and the SoftMax function is used as the activation function at the
output node.

5515
Figure 3. The chromagram of normal voice and dysphonic voice
Figure 4. The architecture of the convolutional neural network
3. RESULTS AND DISCUSSION
To measure the performances of the proposed algorithm, the four machine learning performance
parameters, namely, accuracy, precision, recall, and F1 score [39], [40] are used. Ten simulations were
conducted by using the chromagram of normal and pathological samples. First, the CNN is trained with the
chromagram of 50 normal samples and 50 pathological samples. A five-fold cross-validation method is used
to ensure the accuracy of the training. Once trained, the chromagram of 50 other normal and pathological
samples are used to test the network’s performance. The proposed algorithm’s training, validation, and
testing results are listed in Table 1. This table shows the proposed system achieves an average accuracy of
98.33%, 94.11%, and 84.99% for training, validation, and testing respectively.
The corresponding classification matrix is shown in Table 2. Based on the data presented in Table 2,
it is observed that the proposed system can correctly detect pathological voices with an accuracy of 90.00%.
On the other hand, the system can detect normal voices with an accuracy of 80.00%. This accuracy shows
that the system performs almost equally in detecting normal and dysphonic voices. The classification
matrices also show that the chromagram mistakenly classifies the pathological voices as a normal voice with
a probability of 10.00% only. Also, the chromagram misclassifies the healthy samples as pathological
samples with a probability of 20%. The performances of the proposed algorithm are listed in Table 3. This

 ISSN: 2088-8708
5516
table also shows that the proposed algorithm achieves accuracy, precision, recall, and F1 score of 85.00%,
81.80%, 90.00%, and 85.70%, respectively.
Finally, the performances of the proposed algorithm are compared with other related published
works and the comparison has been presented in Table 4. As listed in the table, the spectrogram audio feature
has been used in [25] and [31] and the authors have achieved an accuracy of 71% in both the works. Table 4
also shows that the proposed system achieved an accuracy of 85% with the chromagram and this accuracy is
higher than that achieved in [25], [31]. Not only that, this accuracy is even close to those of some other
related works presented in [24], [27], [28], [30]. The comparison listed in Table 4 also shows that the
chromagram is a very useful audio feature to detect voice pathology provided a suitable classifier like the
CNN is used. Comparing the results of the proposed system with those of other multiple feature-based
systems, it can be concluded that a single feature like chromagram is sufficient to detect pathological voices
with a high accuracy and hence multiple acoustic features can be avoided to reduce the computation burden
of the pathological voice detection system.
Table 1. Training and testing accuracies for the chromagram
Simulation No. Accuracy (%)
Training Validation Testing
1 95.83 92.00 79.16
2 95.83 90.00 83.33
3 100.00 93.45 87.50
4 100.00 92.33 87.50
5 100.00 96.67 87.50
6 95.83 92.33 79.16
7 100.00 98.00 87.50
8 100.00 96.00 83.33
9 95.83 92.33 87.50
10 100.00 98.00 87.50
Average 98.33 94.11 84.99
Table 2. The classification matrix for the chromagram
Prediction (%)
Actual Normal Pathology
Normal 80.00 20.00
Pathology 10.00 90.00
Table 3. The performance measures for the chromagram
Performance Measures Chromagram (%)
Accuracy 85.00
Precision 81.80
Recall/Sensitivity 90.00
F1 Score 85.70
Table 4. The comparison of the performances
Research works Phonemes Features Tools Accuracy
Wu et al. [31] Vowels, Speech Spectrogram CNN, CDBN 71%
Fang et al. [24] Vowels Mel frequency cepstral
coefficients (MFCC)
SVM, GMM, DNN 94.26% (male)
90.52% (female)
Jun and Kim [25] General voice samples Mel spectrogram DNRNN 71%
Wang and
Cheolwoo [27]
Sustained vowel ‘/a/’ Feature vectors Hidden Markov model (HMM),
GMM, and SVM
95.2%
Sellam,and
Jagadeesan [28]
Tamil phrases Feature vectors SVM, and RBFNN 91% (RBFNN)
83% (SVM)
Sassou [30] Japanese Vowel Higher-order local auto-
correlation (HLAC)
FFNN, and autoregressive hidden
Markov model (AR-HMM)
87.75%
Proposed Method Sustained Vowel ‘/a/’ Chromagram CNN 85.00%
4. CONCLUSION
This paper presents a CNN-based non-invasive pathological voice detection algorithm using
chromagram feature of voice samples. The simulation results suggest that the chromagram is a useful audio
feature that can be used for pathological voice detection algorithms, although chromagram only attracted the

5517
attention of the researchers in music detection. Dysphonia causes a change in the pitch of the sound and
hence affects the chromagram feature more compared to other features available in the literature. Hence, a
higher accuracy is achieved with the chromagram compared to those of other unique features. Other
performance parameters including accuracy, precision, recall, and F1 score, also confirm the effectiveness of
the proposed algorithm. The performances of the proposed algorithm have been compared with those of other
related works available in the literature. This paper also shows that a pathological voice detection system
implemented with the chromagram voice feature and the CNN can outperform some other existing systems
available in the literature. The proposed algorithm discriminates only the dysphonic voices from the normal
voices. However, determining the progression level of voice pathology is another challenging task and this
issue has not been addressed in this work. Other popular machine learning algorithms such as SVM, KNN,
and GMM also need to be investigated in the future to ensure the comparative effectiveness of the proposed
algorithm.
REFERENCES
[1] “Voice disorders,” Johns Hopkins Medicine. https://www.hopkinsmedicine.org/health/conditions-and-diseases/voice-disorders
(accessed Mar. 10, 2022).
[2] D. Taib, M. Tarique, and R. Islam, “Voice feature analysis for early detection of voice disability in children,” in 2018 IEEE
International Symposium on Signal Processing and Information Technology (ISSPIT), Dec. 2018, pp. 12–17, doi:
10.1109/ISSPIT.2018.8642783.
[3] R. Islam, M. Tarique, and E. Abdel-Raheem, “A survey on signal processing based pathological voice detection techniques,”
IEEE Access, vol. 8, pp. 66749–66776, 2020, doi: 10.1109/ACCESS.2020.2985280.
[4] “Voice disorders,” American Speech-Language-Hearing Association (ASHA). https://www.asha.org/practice-portal/clinical-
topics/voice-disorders/ (accessed Mar. 10, 2022).
[5] R. J. Stachler et al., “Clinical practice guideline: hoarseness (dysphonia) (update),” Otolaryngology-Head and Neck Surgery,
vol. 158, Mar. 2018, doi: 10.1177/0194599817751030.
[6] R. H. G. Martins, H. A. do Amaral, E. L. M. Tavares, M. G. Martins, T. M. Gonçalves, and N. H. Dias, “Voice disorders: etiology
and diagnosis,” Journal of Voice, vol. 30, no. 6, Nov. 2016, doi: 10.1016/j.jvoice.2015.09.017.
[7] A. C. Davis, “Vocal cord dysfunction,” Johns Hopkins Medicine. Accessed: Mar. 10, 2022. [Online]. Available:
https://www.hopkinsmedicine.org/health/conditions-and-diseases/vocal-cord-dysfunction
[8] M. M. Johns, R. T. Sataloff, A. L. Merati, and C. A. Rosen, “Article commentary: shortfalls of the American academy of
otolaryngology-head and neck surgery’s clinical practice guideline: hoarseness (dysphonia),” Otolaryngology-Head and Neck
Surgery, vol. 143, no. 2, pp. 175–177, Aug. 2010, doi: 10.1016/j.otohns.2010.05.026.
[9] S. R. Collins, “Direct and indirect laryngoscopy: equipment and techniques,” Respiratory Care, vol. 59, no. 6, pp. 850–864, Jun.
2014, doi: 10.4187/respcare.03033.
[10] D. D. Mehta and R. E. Hillman, “Current role of stroboscopy in laryngeal imaging,” Current Opinion in Otolaryngology and
Head and Neck Surgery, vol. 20, no. 6, pp. 429–436, Dec. 2012, doi: 10.1097/MOO.0b013e3283585f04.
[11] Y. D. Heman-Ackah, S. Mandel, R. Manon-Espaillat, M. M. Abaza, and R. T. Sataloff, “Laryngeal electromyography,”
Otolaryngologic Clinics of North America, vol. 40, no. 5, pp. 1003–1023, Oct. 2007, doi: 10.1016/j.otc.2007.05.007.
[12] A. J. Moshayedi, A. M. Agda, and M. Arabzadeh, “Designing and implementation a simple algorithm considering the maximum
audio frequency of persian vocabulary in order to robot speech control based on arduino,” in Lecture Notes in Electrical
Engineering, Springer Singapore, 2019, pp. 331–346.
[13] R. Islam, E. Abdel-Raheem, and M. Tarique, “Early detection of COVID-19 patients using chromagram features of cough sound
recordings with machine learning algorithms,” in 2021 International Conference on Microelectronics (ICM), Dec. 2021,
pp. 82–85, doi: 10.1109/ICM52667.2021.9664931.
[14] R. Islam, E. Abdel-Raheem, and M. Tarique, “A study of using cough sounds and deep neural networks for the early detection of
Covid-19,” Biomedical Engineering Advances, vol. 3, 100025, Jun. 2022, doi: 10.1016/j.bea.2022.100025.
[15] A. J. Moshayedi, M. S. Hosseini, and F. Rezaei, “WiFi-based massager device with NodeMCU through Arduino interpreter,”
Journal of Simulation and Analysis of Novel Technologies in Mechanical Engineering, vol. 11, no. 1, pp. 73–79, 2018.
[16] O. T. Esfahani and A. J. Moshayedi, “Design and development of Arduino healthcare tracker system for Alzheimer patients,”
International Journal of Innovative Technology and Exploring Engineering (IJITEE), vol. 5, no. 4, pp. 12–16, 2015.
[17] A. Al-Nasheri et al., “Voice pathology detection and classification using auto-correlation and entropy features in different
frequency regions,” IEEE Access, vol. 6, pp. 6961–6974, 2018, doi: 10.1109/ACCESS.2017.2696056.
[18] R. Islam and M. Tarique, “Classifier based early detection of pathological voice,” in 2019 IEEE International Symposium on
Signal Processing and Information Technology (ISSPIT), Dec. 2019, pp. 1–6, doi: 10.1109/ISSPIT47144.2019.9001836.
[19] M. Alhussein and G. Muhammad, “Voice pathology detection using deep learning on mobile healthcare framework,” IEEE
Access, vol. 6, pp. 41034–41041, 2018, doi: 10.1109/ACCESS.2018.2856238.
[20] N. P. Narendra and P. Alku, “Glottal source information for pathological voice detection,” IEEE Access, vol. 8, pp. 67745–67755,
2020, doi: 10.1109/ACCESS.2020.2986171.
[21] R. Islam, E. Abdel-Raheem, and M. Tarique, “A novel pathological voice identification technique through simulated cochlear
implant processing systems,” Applied Sciences, vol. 12, no. 5, 2398, Feb. 2022, doi: 10.3390/app12052398.
[22] P. Harar, J. B. Alonso-Hernandezy, J. Mekyska, Z. Galaz, R. Burget, and Z. Smekal, “Voice pathology detection using deep
learning: a preliminary study,” in 2017 International Conference and Workshop on Bioinspired Intelligence (IWOBI), Jul. 2017,
pp. 1–4, doi: 10.1109/IWOBI.2017.7985525.
[23] P. Murphy, “Development of acoustic analysis techniques for use in the diagnosis of vocal pathology,” Ph.D. Dissertation,
School of Physical Sciences, Dublin City University, Dublin, 1997, [Online]. Available:
https://doras.dcu.ie/19122/1/Peter_Murphy_20130620152522.pdf
[24] S.-H. Fang et al., “Detection of pathological voice using cepstrum vectors: a deep learning approach,” Journal of Voice, vol. 33,
no. 5, pp. 634–641, Sep. 2019, doi: 10.1016/j.jvoice.2018.02.003.
[25] T. J. Jun and D. Kim, “Pathological voice disorders classification from acoustic waveforms.” 2018. [Online]. Available:
https://mac.kaist.ac.kr/~juhan/gct634/2018/finals/pathological_voice_disorders_classification_from_acoustic_waveforms_report.p

 ISSN: 2088-8708
5518
df
[26] V. Srinivasan, V. Ramalingam, and P. Arulmozhi, “Artificial neural network based pathological voice classification using MFCC
features,” International Journal of Science, Environment, vol. 3, no. 1, pp. 291–302, 2014.
[27] J. Wang and C. Jo, “Vocal folds disorder detection using pattern recognition methods,” in 2007 29th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society, Aug. 2007, pp. 3253–3256, doi:
10.1109/IEMBS.2007.4353023.
[28] V. Sellam and J. Jagadeesan, “Classification of normal and pathological voice using SVM and RBFNN,” Journal of Signal and
Information Processing, vol. 05, no. 01, pp. 1–7, 2014, doi: 10.4236/jsip.2014.51001.
[29] M. Chopra, K. Khieu, and T. Liu, “Classification and recognition of stuttered speech.” Stanford University, [Online]. Available:
https://web.stanford.edu/class/cs224s/project/reports_2017/Manu_Chopra.pdf.
[30] A. Sasou, “Automatic identification of pathological voice quality based on the GRBAS categorization,” in 2017 Asia-Pacific
Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Dec. 2017, pp. 1243–1247, doi:
10.1109/APSIPA.2017.8282229.
[31] H. Wu, J. Soraghan, A. Lowit, and G. Di-Caterina, “A deep learning method for pathological voice detection using convolutional
deep belief networks,” in Interspeech 2018, Sep. 2018, pp. 446–450, doi: 10.21437/Interspeech.2018-1351.
[32] M. A. Mohammed et al., “Voice pathology detection and classification using convolutional neural network model,” Applied
Sciences, vol. 10, no. 11, 3723, May 2020, doi: 10.3390/app10113723.
[33] L. Verde, G. De Pietro, M. Alrashoud, A. Ghoneim, K. N. Al-Mutib, and G. Sannino, “Dysphonia detection index (DDI): a new
multi-parametric marker to evaluate voice quality,” IEEE Access, vol. 7, pp. 55689–55697, 2019, doi:
10.1109/ACCESS.2019.2913444.
[34] H. Ding, Z. Gu, P. Dai, Z. Zhou, L. Wang, and X. Wu, “Deep connected attention (DCA) ResNet for robust voice pathology
detection and classification,” Biomedical Signal Processing and Control, vol. 70, 102973, Sep. 2021, doi:
10.1016/j.bspc.2021.102973.
[35] M. Putzer and W. J. Barry, “Saarbrücken voice database.” Accessed: Mar. 10, 2022. [Online]. Available: http://stimmdb.coli.uni-
saarland.de/index.php4#target.
[36] D.L. Wang and G. J. Brown, "Fundamentals of computational auditory scene analysis," in Computational Auditory Scene
Analysis: Principles, Algorithms, and Applications, IEEE, 2006, pp.1-44, doi: 10.1109/9780470043387.ch1.
[37] R. D. Patterson, K. Robinson, J. Holdsworth, D. McKeown, C. Zhang, and M. Allerhand, “Complex sounds and auditory images,”
in Auditory Physiology and Perception, Elsevier, 1992, pp. 429–446.
[38] P. Kim, “Convolutional neural network,” in MATLAB Deep Learning, Berkeley, CA: Apress, 2017, pp. 121–147.
[39] R. M. Rangayyan, “Pattern classification and diagnostic decision,” in Biomedical Signal Analysis, Hoboken, NJ, USA: John
Wiley & Sons, Inc., 2015, pp. 571–632.
[40] Y. Jiao and P. Du, “Performance measures in evaluating machine learning based bioinformatics predictors for classifications,”
Quantitative Biology, vol. 4, no. 4, pp. 320–330, Dec. 2016, doi: 10.1007/s40484-016-0081-2.
BIOGRAPHIES OF AUTHORS
Rumana Islam completed her Master of Science (MSc) in Biomedical
Engineering from Wayne State University, Michigan in 2004. Currently, she is pursuing her
Doctor of Philosophy (Ph.D.) degree in the Department of Electrical and Computer
Engineering, University of Windsor, Canada. She has professional experience in academy
and industry as well. She served as a professional Engineer in SIEMENS and Bangladesh
Telegraph and Telephone Board (BTTB) in Bangladesh. She served as an Adjunct Faculty
member at East West University of Bangladesh. She worked as a Lecturer in the Department
of Electrical and Electronic Engineering, American International University, Bangladesh
(AIUB) from 2008 to 2011. Currently, she is working as an Adjunct Faculty member in the
Department of Electrical and Computer Engineering at the University of Science and
Technology of Fujairah (USTF), UAE. She has published several research papers in highly
reputed journals and conference proceedings. Her research interests include biomedical
engineering, biomedical signal processing, biomedical diagnostic tools using machine
learning, digital signal processing, and biomedical image processing. She can be contacted at
email: islamq@uwindsor.ca.
Mohammed Tarique completed his Doctor of Philosophy (Ph.D.) degree in
Electrical Engineering from the University of Windsor, Canada in 2007. He had worked as an
Assistant Professor in the Department of Electrical Engineering, American International
University-Bangladesh from 2007-2011. He joined Ajman University in 2011. Currently, he
is working as an Associate Professor of Electrical Engineering at the University of Science
and Technology of Fujairah (USTF). His research interests are in wireless communication,
mobile ad hoc networks, digital communication, digital signal processing, and biomedical
signal processing. He has published 40 research papers in the top-ranked journals in the
above-mentioned fields. He also published 15 research papers in the proceedings of highly
reputed international conferences. He also had worked as a Senior Engineer in Beximco
Pharmaceuticals Limited, Bangladesh from 1993 to 1999. He also completed Master of
Business Administration (MBA) from the Institute of Business Administration (IBA),
University of Dhaka, Bangladesh. He can be contacted at email: m.tarique@ustf.ac.ae.

A novel convolutional neural network based dysphonic voice detection algorithm using chromagram

More Related Content

Similar to A novel convolutional neural network based dysphonic voice detection algorithm using chromagram (20)

More from IJECEIAES (20)

Recently uploaded (20)

A novel convolutional neural network based dysphonic voice detection algorithm using chromagram