IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 4, December 2023, pp. 1901~1908

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i4.pp1901-1908  1901

Improving Indonesian multietnics speaker recognition using

pitch shifting data augmentation

Kristiawan Nugroho1, Isworo Nugroho1, De Rosal Ignatius Moses Setiadi2, Omar Farooq3
1
Department of Information Technology and Industry, Universitas Stikubank, Semarang, Indonesia
2
Department of Computer Science, Universitas Dian Nuswantoro, Semarang, Indonesia
3
Department of Electronics Engineering, Z. H. College of Engg and Technology, A. M. U, Aligarh, India

Article Info ABSTRACT

Article history: Speaker recognition to recognize multiethnic speakers is an interesting
research topic. Various studies involving many ethnicities require the right
Received Jan 25, 2023 approach to achieve optimal model performance. The deep learning approach
Revised Mar 15, 2023 has been used in speaker recognition research involving many classes to
Accepted Mar 27, 2023 achieve high accuracy results with promising results. However, multi-class
and imbalanced datasets are still obstacles encountered in various studies
using the deep learning method which cause overfitting and decreased
Keywords: accuracy. Data augmentation is an approach model used in overcoming the
problem of small amounts of data and multiclass problems. This approach can
Data augmentation improve the quality of research data according to the method applied. This
Deep learning study proposes a data augmentation method using pitch shifting with a deep
Pitch shifting neural network called pitch shifting data augmentation deep neural network
Speaker recognition (PSDA-DNN) to identify multiethnic Indonesian speakers. The results of the
research that has been done prove that the PSDA-DNN approach is the best
method in multi-ethnic speaker recognition where the accuracy reaches
99.27% and the precision, recall, F1 score is 97.60%.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Kristiawan Nugroho
Department of Information Technology and Industry, Universitas Stikubank Semarang
Jl. Tri Lomba Juang, Semarang, Indonesia
Email: [email protected]

1. INTRODUCTION
Speaker recognition is one of the challenging research fields. Various kinds of problems need to be
solved to produce the discovery of new theories that can make a positive contribution to human life. Research
results in the field of speaker recognition have proven to have led to various forms of new technology applied
to voice authentication, surveillance speaker recognition, forensic speaker recognition, security, and multi
speaker tracking. Tech giant companies such as Google, Microsoft, and Amazon have also taken advantage of
this technology, including Google Voice technology, Apple Siri, and Amazon's Alexa.
The application of this technology is proven to be able to help human work, such as in the field of
security and authentication in smart homes to help people with disabilities to recognize sound patterns and
images. Initially, research in the field of speaker recognition used classical methods in machine learning,
gaussian mixture model (GMM), such as studies carried out by Motlicek et al. [1] and Veena and Mathew [2].
In speaker recognition, the hidden markov model (HMM) strategy is also utilized, as demonstrated by
Maghsoodi et al. [3], Hussein, et al. [4] and the support vector machine (SVM) approach on research that has
been done by Chaunan et al. [5].
However, along with the growth of data that is getting bigger and the complexity of the problems
faced today, researchers are starting to use the deep learning method in speaker recognition research. Deep

Journal homepage: http://ijai.iaescore.com

1902  ISSN: 2252-8938

learning (DL) is an approach developed from the neural network algorithm. This method continues to be
developed by researchers to solve various problems in machine learning. DL has advantages in terms of the
ability to handle computational processes that involve very large data. In addition, DL is also able to process
data representations of various forms such as text, images, and sound, which makes it a good ability to process
information in the form of multi modals so that this method outperforms the previous machine learning method,
especially used in the field of computer vision.
Various DL methods are used in voice signal processing research using the deep neural networks
(DNN) approach as on research that has been done by Guo et al. [6] combined with I-Vector in research on
short speech. This study improves the equal error rate (EER) to 26.47%. In other research Mohan and Patil [7]
using self organizing map (SOM) and latent dirichlet allocation (LDA) succeeded in increasing crop prediction
accuracy by 7-23%. DNN is also used by Saleem and Khattak [8] in research on speech separation, which
produces promising performance. However, behind the advantages of DL, this method also has several
weaknesses, including higher accuracy performance requiring large data sets, overfitting, and computational
process efforts that require large resources. In addition to the need for large data volumes, research in the field
of modern classification also faces the problem of multiple classes when it comes to processing an explosion
of features and limited data.
In several studies regarding speaker recognition multi ethnic, the problem of limited data is an initial
challenge in conducting research such as the study conducted by Hanifa et al. [9], which only had 62 recorded
data of Malaysian speakers, and the study of Cole [10] to identify speakers in South East England with 227
speakers. To solve the problem of limited data, various approaches are used, among others, by using ata
augmentation (DA). DA is a method of expanding the quantity of data and has proven to be effective in
conducting training on neural network. In speech signal processing, DA is proven to increase accuracy in audio
classification using DL [11]. In the field of speech signal processing, several DA methods that are often used
are adding white noise (AWN), time stretching (TS), pitch shifting (PS), mixupand speech augment. Some of
these DA approaches are proven to be able to increase the quantity of research data so that research using DL
which requires relatively large amounts of data can be done well so as to achieve a high level of accuracy.
Currently, in machine learning, the problems faced by researchers in an effort to increase the accuracy
of speaker recognition include the number of classes that must be classified in the imbalance dataset. One of
the problems encountered in machine learning classification is unbalanced multiclass which will cause model
inaccuracies in predicting data. This problem can cause prediction errors in machine learning algorithms, so it
needs to be resolved immediately. In research conducted by Khan et al. [12] and Mi et al. [13] the DA approach
used the generative adversarial network (GAN) method to solve multi-class problems which achieved an
accuracy increase of 6.4%. In several studies related to the implementation of DA in speech signal processing,
various methods used include AWN, such on research that has been done by Morales et al. [14], which seeks
to increase speaker recognition accuracy by adding noise effects.
Jacques and Roebel [15] also used the approach to adding noise in the research conducted. TS is also
used by several researchers in speech signal processing research, TS functions to change the speed of audio
signal duration as on research that has been done by Sasaki et al. [16] and Aguiar et al. [17] to classify music
genres with convolutional neural network (CNN). However, the AWN and TS methods used still cannot
produce high accuracy for speech recognition because they only achieve an accuracy level of around 70% to
80%. Another approach used in voice-based augmentation data is PS. PS method is also used in several studies
conducted by Morbale and Navale [18] in processing audio files, Rai and Barkana [19] in processing musical
instruments and Ye et al. [20] with CNN using UME and TIMIT datasets. In research conducted by Ye the
Pitch Shifting method has achieved an accuracy rate above 90% with the highest accuracy of 98.72%.
This paper aims to improve the performance of the multi-ethnic speaker recognition model with a
pitch shifting method based on deep neural networks and consists of several parts, namely the Introduction
section in Chapter 1, which describes the research problems and other research that has been carried out and
the relationship with several other studies. Chapter 2 is a literature review that contains is a section that contains
the study of related topics. In Chapter 3, proposed methodology contains the proposed model in solving the
problem. Chapter 4 is a results and discussion that contains experiment results from the research that has been
done. The final part is Chapter 5, namely conclusion, which contains several conclusions and solutions for the
research carried out.

2. METHOD
This study proposed a pitch shifting deep neural network (PSDA-DNN) which has the best
performance when implemented in voice signal processing supported by MFCC as a feature extraction method
and a DNN that processes multiethnic speaker recognition classification. This research begins with processing
the dataset of multiethnic speakers followed by the preprocessing process. Data augmentation is a step that

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1901-1908

Int J Artif Intell ISSN: 2252-8938  1903

provides a solution to the limited dataset of multiethnic speakers. The dataset is then extracted using the mel
frequency cepstral coefficient (MFCC) approach and then the result of feature extraction is included in the 7
Layer DNN architecture. As the last step, measuring the performance results of the proposed framework, among
others, uses measures of accuracy, precision, recall, and F1. The proposed method can be seen in Figure 1.

Figure. 1 Proposed method

2.1. Datasets and preprocessing

Research on multi-ethnic speaker recognition uses the sample for recognizing ethnic speakers dataset
from Indonesia taken from Youtube 301 language in Indonesia [21]. The dataset was compiled using 70 ethnic
male speakers from hundreds of ethnic groups in Indonesia. Sound processing using Adobe Audition CS6, the
voice of tribal speakers is carried out by a sampling process by taking 10 tribal speakers each with a duration
of 1 second. The sampling process uses a standard sample rate (SR) of 44.100 Hz with 32-bit bit depth, mono
32-bit Floating Point. The next stage is the reprocessing process, which is an important process in data mining
processing. Preprocessing is a step that needs to be done in data mining to get good data quality to reduce
processing time and get the desired results. This study uses the Adobe Audition CS6 application in
preprocessing data with noise reduction facilities to remove noise in the speaker's speech data.

2.2. Data augmentation

DA is one method that is often used in increasing the quantity of dataset needed in research. DA is an
approach that aims to increase the size of the data quantity and is a powerful technique used in the field of data
mining and data processing for regression and classification purposes. PS is a DA method in sound signal
processing by raising or lowering the original voice pitch in audio without affecting the long duration of the
recorded sound. PS is used in this study because it has the advantage that the overall spectral envelope does
not change so that it can achieve high-quality output The process in the PS approach can be seen in Figure 2.
In this study, PS results will be compared with 2 other DA methods, namely AWN and TS which are widely
implemented in various research fields such as sound recording, music production, music learning, and foreign
languages. The original speaker's voice signal is processed using the Pitch Shifting approach. The results of
processing the voice signal before and after using the PS method can be seen in Figure 2.

2.3. Feature extraction and deep neural network seven layer

2.3.1. Mel-frequency cepstral coefficient
MFCC is a feature extraction method used in this study. MFCC is a robust approach used in speech
recognition. The speech signal of the ethnic speaker consisting of 700 wav voices from 70 Ethnic was extracted
using the python application with MFCC settings of frame 900 lengths, 25 frame shifts 10, window type
hamming, preemphasis coefficient 0.97, number of cepstral coefficient 13 and number of lifters 22. In the
MFCC approach, the voice signal will be processed through the following steps:
i) Preemphasis
Is a process that will be carried out after the sound sampling process, this process serves to reduce noise at
the sound source. The preemphasis process in the time domain can be formulated as:

𝑦(𝑛) = 𝑥 (𝑛) − 𝑎𝑥 (𝑛 − 1) (1)

A indicates the constant which is at 0.9 < a < 1.0

Improving Indonesian multietnics speaker recognition using pitch shifting data… (Kristiawan Nugroho)
1904  ISSN: 2252-8938

Figure 2. Original and pitch shifting speech signal

ii) Frame blocking

The speech signal will then enter the Frame Blocking process which functions to divide the sound into
several parts of the frame.
iii) Windowing
Windowing is a step to analyze a long signal by taking the right part to be processed in the next stage. If
the window is defined as w(n), 0 < n < N – 1 where N is the number of samples in each frame, then the
windowing signal can be formulated as:

𝑦1(𝑛) = 𝑥1(𝑛) 𝑤 (𝑛) , 0 ≤ 𝑛 ≤ 𝑁 − 1 (2)

iv) Fast fourier transform (FFT)

Fourier transform is used to convert the time series of signals in the form of a limited time domain into a
frequency spectrum, while FFT is part of a fast Discrete Fourier Transform (DFT) algorithm that converts
frames into N samples starting from the time domain to the frequency domain. The result of this processing
is referred as Cepstrum which is formulated as:

𝑥(𝑛) = ∑𝑁−1
𝑘=0 𝑥𝑘 𝑒
−2𝜋𝑗𝑘𝑛/𝑁
(3)

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1901-1908

Int J Artif Intell ISSN: 2252-8938  1905

Where n = 0.1,2,..,N-1 and j = sqrt-1 while X[n] is an n-frecuency form resulting from the Fourier
Transform mechanism.
v) Mel-frequency wrapping
In this step the existing FFT signals are grouped in a triangular filter which aims to multiply the FFT value
with the appropriate filter gain which then the results will be summed. The wrapping process into a signal
in the frequency domain can be formulated as:

𝑥𝑖 = log10 (∑𝑁−1
𝑘=0 |𝑥(𝑘) 𝐻𝑖 (𝑘) ) (4)

Where i = 1, 2, 3, .., M, M is the sum of the triangle filters and Hi(k) is the value for the triangular i-filter
for the acoustic frequency k.
vi) Cepstrum
In order for the signal to be heard by humans, it is necessary to convert the signal into a time domain using
a discrete cosine transform (DCT). The final result of this process is referred to as Mel Frequency Cepstral
Coefficients which is formulated as:

𝑐𝑗 = ∑𝐾
𝑗=1 𝑥𝑗 cos ( 𝑗(𝑖−1)/2 𝜋 ) (5)
𝐾

Cj shows the MFCC coefficient. Xj is the strength of the Mel Frequency spectrum, j = 1, 2, 3,.., K is the
expected coefficient and M indicates the number of filters.

2.3.2. Deep neural network seven layer

One of the finest DL techniques is the DNN. DNN has the advantage of building more accurate
models. In this work used DNN 7 layers with architectural network as in Table 1. The DNN architecture used
in this study consists of seven layers consisting of dense or also called fully connected layers, indicating that
the layer in which there are neurons connected to neurons in the previous layer. Layer 1 consists of 193 nodes
which is an input layer that shows 193 features generated from the extracted features. Between layers is given
a dropout function which is a technique used to solve overfitting problems and prediction problems in large
neural networks. Layers two to seven use half of the number of nodes in the previous layer in order to reduce
the complexity of calculations on each layer.

Table 1. DNN7L architecture

Layer (type) Output shape Layer (type) Output shape
Dense 1 (dense) (None,193) Dense 5 (dense) (None,50)
Dense 2 (dense) (None,400) Dropout_4 (Dropout) (None,50)
Dropout_1 (Dropout) (None,400) Dense 6 (dense) (None,25)
Dense 3 (dense) (None,200) Dropout_5 (Dropout) (None,25)
Dropout_2 (Dropout) (None,200) Dense 7 (dense) (None,15)
Dense 4 (dense) (None,100) Dropout_6 (Dropout) (None,15)
Dropout_3 (Dropout) (None,100) Dense 8 (dense) (None,8)

2.4. Evaluation
The end result of the process of Indonesian ethnic speakers recognition is an evaluation process by
measuring the performance of the proposed model by evaluating the level of accuracy, precision, recall, and
F1 measure. Accuracy is the ratio of the number of cases that were predicted with correct answers compared
to the total number of cases, while recall is the ratio of the number of positive cases that were predicted correctly
to the number of positive cases that were predicted. Accuracy, precision, recall and F1-score measure can be
calculated using the (6) to (9), respectively.
𝑇𝑃+𝑇𝑁
𝑥 100% (6)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

𝑇𝑃
𝑥 100% (7)
(𝑇𝑃+𝐹𝑃)

𝑇𝑃
𝑥 100% (8)
(𝑇𝑃+𝐹𝑁)

2 𝑋 (𝑅𝑒𝑐𝑎𝑙𝑙 𝑋 𝑃𝑟𝑒𝑐𝑖𝑠𝑠𝑖𝑜𝑛)
(9)
(𝑅𝑒𝑐𝑎𝑙𝑙+ 𝑃𝑟𝑒𝑐𝑖𝑠𝑠𝑖𝑜𝑛)

Improving Indonesian multietnics speaker recognition using pitch shifting data… (Kristiawan Nugroho)
1906  ISSN: 2252-8938

Where TP, TN, FP, and FN stand for true positive, true negative, false positive, and false negative, respectively.
Performance measurement in this study also uses recall which is the ratio of true positive cases that are
predicted to be positive.

3. RESULTS AND DISCUSSION

Dataset on multi-ethnic speakers in Indonesia were tested using DA approach, namely AWN, PS
and TS then trained using DNN using a split ratio of 70:30, 80:20 and 90:10. The test results show that the
PSDA-DNN has better performance than adding white noise data augmentation deep neural network
(AWNDA- DNN), and time stretching data augmentation deep neural network (TSDA-DNN) methods. The
following are the results of the PSDA-DNN model performance testing process using a 70:30 split ratio as
depicted in Figure 3.
Testing on the 70:30 split ratio resulted in an accuracy level of 98.55%, precision, recall and F1
measure each of 94.37%. This performance result shows that PSDA-DNN will be more robust when used on
larger data. Classification using various machine learning methods using many classes is not easy and cause
various problems in learning [22]. An appropriate approach is needed in managing the dataset. The following
is a comparison of the number of classes processed between the various types of research shown in Table 2.

Figure 3. PSDA-DNN performance

Table 2. Comparison with another datasets

Methods Datasets ∑Class Acc (%)
SVM [9] Ethnicity of Malaysian dataset 4 57.7
CNN [23] Urdu speakers 4 87.5
Deep Belief Network (DBN) [24] Accented spoken English corpus 6 90.2
DNN [25] TITML-IDN, OpenSLR 4 98.9
PSDA-DNN Indonesian multietnics speakers 42 99.2

Based on the comparison results presented in Table 2, even though it has more classes, the proposed
model, namely PSDA-DNN, produces better model performance than other machine learning methods with an
accuracy of up to 99.2%. The achievement of a high level of performance at PSDA-DNN was due to a good
preprocessing process on the speech signal dataset which was then carried out by data augmentation process
with PS and proper classification using the DNN 7-layer approach. The PSDA-DNN method also produces the
most effective performance in comparison to various other techniques using the Indonesian Multiethnics
Speakers dataset. The results of the comparison of these methods can be seen in Table 3.

Table 3. Comparison with another methods

Methods Accuracy
K-Nearest Neighbor 92%
Random-Forest 81%
DNN 98.4%
PSDA-DNN (Ours) 99.2%

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1901-1908

Int J Artif Intell ISSN: 2252-8938  1907

The results of this study will be compared with several methods with classical machine learning and
deep learning approaches. Table 3 shows that the PSDA-DNN method has better performance compared to
other methods. In several comparisons of the performance of the methods as previously presented, it can be
concluded that the proposed method has the best performance.

4. CONCLUSION
Study in the area of speaker identification is an interesting topic and challenges researchers around
the world to work hard to make new scientific contributions, including research on Indonesian multiethnic
speaker recognition. The DL method is an approach that has been chosen, especially for processing large
amounts of data, including sound signal processing. However, the problem of multiple classes and data
imbalanced causes low accuracy because the model performance is not optimal. The PS Approach in Data
Augmentation is a solution in increasing the quantity of data and as a solution for multiple class classification
problems. Obtaining a high model accuracy performance of 99.27% through the proposed model, namely
PS-DNN is one of the solutions to overcome the problem of multiple classes and imbalanced datasets in
machine learning. This study proposes the PSDA-DNN approach which is a multi-ethnic speaker recognition
method that uses the PSDA technique which is supported by the MFCC and DNN methods in processing speech
signals. The research results show that PSDA-DNN has better performance compared to other approaches such
as AWN and TS which are also DNN-based in processing speech signals. The PSDA-DNN approach produces
an average level of accuracy of 99.27%, precision, recall, and F1 measure of 97.60%.

REFERENCES
[1] P. Motlicek, S. Dey, S. Madikeri, and L. Burget, “Employment of subspace gaussian mixture models in speaker recognition,” in
2015 IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP, Apr. 2015, pp. 4445–4449,
doi: 10.1109/icassp.2015.7178811.
[2] K. V Veena and D. Mathew, “Speaker identification and verification of noisy speech using multitaper MFCC and gaussian mixture
models,” Dec. 2015, doi: 10.1109/picc.2015.7455806.
[3] N. Maghsoodi, H. Sameti, H. Zeinali, and T. Stafylakis, “Speaker recognition with random digit strings using uncertainty normalized
HMM-based i-vectors,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 11, pp. 1815–1825,
Nov. 2019, doi: 10.1109/taslp.2019.2928143.
[4] J. S. Hussein, A. A. Salman, and T. R. Saeed, “Arabic speaker recognition using HMM,” Indonesian Journal of Electrical
Engineering and Computer Science, vol. 23, no. 2, pp. 1212–1218, Aug. 2021, doi: 10.11591/ijeecs.v23.i2.pp1212-1218.
[5] N. Chauhan, T. Isshiki, and D. Li, “Speaker recognition using LPC, MFCC, ZCR features with ANN and SVM classifier for large
input database,” in 2019 IEEE 4th International Conference on Computer and Communication Systems ICCCS, Feb. 2019,
pp. 130–133, doi: 10.1109/ccoms.2019.8821751.
[6] J. Guo et al., “Deep neural network based i-vector mapping for speaker verification using short utterances,” Speech Communication,
vol. 105, pp. 92–102, Dec. 2018, doi: 10.1016/j.specom.2018.10.004.
[7] P. Mohan and Kiran Patil, “Deep learning based weighted SOM to forecast weather and crop prediction for agriculture application,”
International Journal of Intelligent Engineering and Systems, vol. 11, no. 4, pp. 167–176, Aug. 2018,
doi: 10.22266/ijies2018.0831.17.
[8] N. Saleem and M. I. Khattak, “Deep neural networks based binary classification for single channel speaker independent multi-talker
speech separation,” Applied Acoustics, vol. 167, p. 107385, Oct. 2020, doi: 10.1016/j.apacoust.2020.107385.
[9] R. M. Hanifa, K. Isa, and S. Mohamad, “Speaker ethnic identification for continuous speech in Malay language using pitch and
MFCC,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 19, no. 1, pp. 207–214, Jul. 2020,
doi: 10.11591/ijeecs.v19.i1.pp207-214.
[10] A. Cole, “Identifications of speaker ethnicity in South-East England: multicultural London English as a divisible perceptual variety,”
in Proceedings of the LREC 2020 Workshop on Citizen Linguistics in Language Resource Development, 2020, pp. 49–57.
[11] L. Nanni, G. Maguolo, and M. Paci, “Data augmentation approaches for improving animal audio classification,” Ecological
Informatics, vol. 57, p. 101084, May 2020, doi: 10.1016/j.ecoinf.2020.101084.
[12] M. H.-M. Khan et al., “Multi-class skin problem classification using deep generative adversarial network (DGAN),” Computational
Intelligence and Neuroscience, vol. 2022, pp. 1–13, Mar. 2022, doi: 10.1155/2022/1797471.
[13] Q. Mi, Y. Hao, M. Wu, and L. Ou, “An enhanced data augmentation approach to support multi-class code readability classification,”
in Proceedings of the Inernational Conference on Software Engineering and Knowledge Engineering SEKE, Jul. 2022, pp. 48–53,
doi: 10.18293/SEKE2022-130.
[14] N. Morales, L. Gu, and Y. Gao, “Adding noise to improve noise robustness in speech recognition,” Jan. 2007,
doi: 10.21437/interspeech.2007-335.
[15] C. Jacques and A. Roebel, “Data augmentation for drum transcription with convolutional neural networks,” Sep. 2019,
doi: 10.23919/eusipco.2019.8902980.
[16] T. Sasaki et al., “Time stretching: illusory lengthening of filled auditory durations,” Attention, Perception, Psychophys, vol. 72,
no. 5, pp. 1404–1421, Jul. 2010, doi: 10.3758/app.72.5.1404.
[17] R. L. Aguiar, Y. M. G. Costa, and C. N. Silla, “Exploring data augmentation to improve music genre classification with convNets,”
in 2018 International Joint Conference on Neural Networks IJCNN, Jul. 2018, pp. 1–8, doi: 10.1109/ijcnn.2018.8489166.
[18] P. R. Morbale and M. Navale, “Design and implementation of real time audio pitch shifting on FPGA,” International Journal of
Innovative Trends in Engineering, vol. 4, no. 2, pp. 81–88, 2015.
[19] A. Rai and B. D. Barkana, “Analysis of three pitch-shifting algorithms for different musical instruments,” in 2019 IEEE Long Island
Systems, Applications and Technology Conference LISAT, May 2019, pp. 1–6, doi: 10.1109/lisat.2019.8817334.
[20] Y. Ye, L. Lao, D. Yan, and R. Wang, “Identification of weakly pitch-shifted voice based on convolutional neural network,”
International Journal of Digital Multimedia Broadcasting, vol. 2020, pp. 1–10, Jan. 2020, doi: 10.1155/2020/8927031.

Improving Indonesian multietnics speaker recognition using pitch shifting data… (Kristiawan Nugroho)
1908  ISSN: 2252-8938

[21] “301 languages in Indonesia-regional dialects #2 (In Indonesia: 301 languages in Indonesia-bahasa Logat Dialek Daerah #2).”
Indonesia Ideas, 2018, [Online]. Available: https://www.youtube.com/watch?v=FkwXbCY1rWg.
[22] Y. Xue and M. Hauskrecht, “Active learning of multi-class classification models from ordered class sets,” Proceedings of the AAAI
Conference on Artificial Intelligence, vol. 33, no. 01, pp. 5589–5596, Jul. 2019, doi: 10.1609/aaai.v33i01.33015589.
[23] A. Ashar, M. S. Bhatti, and U. Mushtaq, “Speaker identification using a hybrid CNN-MFCC approach,” Mar. 2020,
doi: 10.1109/icetst49965.2020.9080730.
[24] R. Upadhyay and S. Lui, “Foreign English accent classification using deep belief networks,” in 2018 IEEE 12th International
Conference on Semantic Computing ICSC, Jan. 2018, pp. 290–293, doi: 10.1109/icsc.2018.00053.
[25] K. Azizah, M. Adriani, and W. Jatmiko, “Hierarchical transfer learning for multilingual, multi-speaker, and style transfer DNN-
based TTS on low-resource languages,” IEEE Access, vol. 8, pp. 179798–179812, 2020, doi: 10.1109/access.2020.3027619.

ACKNOWLEDGEMENTS
We would like thank to the Universitas Stikubank for granted funding through scientific publication
incentives.

BIOGRAPHIES OF AUTHORS

Kristiawan Nugroho works as a lecturer at the faculty of information

technology and industry, Universitas Stikubank. He obtained a bachelor's degree in 2001 in
the information systems department, Faculty of Computer Science, Universitas Dian
Nuswantoro Semarang, then in 2007 He graduated from Universitas Dian Nuswantoro with
a master's degree in informatics engineering. He also obtained Doctoral degree in computer
science with a concentration in Machine Learning and Artificial Intelligence in 2022 at Dian
Nuswantoro University Semarang. He has conducted various researches in machine learning,
speech recognition and sentiment analysis. He can be contacted via email
[email protected].

Isworo Nugroho works as a lecturer at the faculty of information technology

and industry, Universitas Stikubank. He obtained a bachelor's degree in 2001 in the
management department, Faculty of Economis, Stikubank University, then in 2003 he
obtained a Master's degree in computer science, Gajah Mada University. He has conducted
various researches in text processing, data mining and statistical science. He can be contacted
via email [email protected].

De Rosal Ignatius Moses Setiadi a Bachelor of Science in Informatics

Engineering from Universitas Soegijaprana, Indonesia, and a Master of Science in
Informatics Engineering from Universitas Dian Nuswantoro Semarang, both in 2012. He is
presently a lecturer and researcher at the Faculty of Computer Science at Universitas Dian
Nuswantoro in Semarang, Indonesia. He has written more than 138 peer-reviewed journal
and conference publications that Scopus has indexed. His areas of interest in study include
machine learning, cryptography, image steganography, and watermarking. His email address
is [email protected].

Prof. Omar Farooq Omar Farooq joined the Department of Electronics

Engineering, AMU Aligarh as Lecturer in 1992 and is currently working as a professor. He
was awarded Commonwealth Scholar from 1999-2002 towards PhD at Loughborough
University, UK, and a one-year postdoctoral fellowship with the UKIERI in 2007-2008. With
a focus on speech recognition, signal processing is his broad area of study interest. He has
approximately 250 publications published by him or with him in reputable academic journals
and conference proceedings, and he has helped 9 scholars complete their PhDs. He is a Senior
Member, Institute of Electrical and Electronics Engineers, (IEEE, USA). He can be contacted
via email [email protected].

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1901-1908