Volume 5, Issue 10, October – 2020 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Gender Detection by Voice Using Deep Learning

Mia Mutiany Iwa Ovyawan Herlistiono
Department of Informatics Department of Informatics
WidyatamaUniversity Widyatama University
Bandung, Indonesia Bandung, Indonesia

Abstract:- The recognition of gender voices as an MFCC chose feature extraction because it is a fairly
important part of answering certain voices. To good feature extraction method for noise reduction which
distinguish gender from sound signals, sound techniques requires a fast, easy, and complete processing time.
have defined the gender-relevant features (male or Meanwhile, the classification of votes is based on gender
female) of these sound signals. In this study, we used using a machine support vector (SVM). [3]
various models to improve accuracy, one of which was
by using deep learning with the voice gender DNN II. RELATED WORK
method. This noise reduction uses the extraction feature
of the Mel Frequency Cepstral Coefficient (MFCC), then Several studies discussdetecting voices based on
the sound classification uses SVM. By using a separation gender. Previous research conducted by Martin and Joensuu
ratio of 80% for training data and 20% for testing data. who developed speech recognition using the detected GMM
The results showed that using DNN for voice recognition and FFT features gave the best results for the classification
was better and pairing with the SVM algorithm obtained level[4][5].
an accurate result of 0.97%.
S.LYuan[6] developed a voice recognition system and
Keywords:- Voice Recognition, Deep Neural Network, Deep detected gender based on voice using Deep Learning with
Learning, MFCC, SVM. the Deep Neural Network (DNN) algorithm resulting in a
Word Error Rate (WER) in speech recognition which
I. INTRODUCTION showed less than optimal results.

Voice recognition is one of the important researches Also, Lee and Kwak [7] used DNN and two
that are currently widely used from a variety of applications, classifications in detecting sounds based on gender. The two
such as security systems, authentication, and so on. Voice classifications are SVM and decision tree (DT). In his
recognition must have high performance, to be able to research, feature extraction used by MFCC to identify
improve speech recognition performance, one of which is by gender voices resulted in fairly good accuracy.
adding a gender classification procedure. With this gender
classification, the problem space in speech recognition can III. DEEP LEARNING
be limited only based on predetermined gender[1].
Deep learning is a method that is often used in the
Voice data is divided into training data and testing data field of machine learning based on the Network Artificial
by classifying gender into two categories, namely male and Neural (ANN) principle model. Deep learning can solve
female. Male and female voices have their characteristics problems with large datasets such as image recognition, text
due to different resonances in the throat[2]. By processing detection, speech recognition, audio, etc. Because there are
sound signals, these characteristics will be obtained in a techniques for using feature extraction from training data
form that can be recognized by a computer. With especially for speech recognition. Artificial Neural Network,
thesecharacteristics, the computer can identify gender a method that adds a hidden layer, this deep learning can be
through sound signals. Therefore, we need a learning started with the input layer (voice recording), which can
algorithm that can help humans detect sounds based on then be processed in the form of a signal that is
gender. The deep learning algorithm can support this interconnected between nodes with each other in processing
detection because it can predict more accurately and quickly data and ultimately through the output to accuracy.
for speech recognition.
One of the deep learning algorithms is called the Deep
Deep Learning has excellent models for detections Neural Network (DNN). DNN is one of the developments of
such as image recognition, emotional recognition, and the Artificial Neural Network. The DNN method is capable
speech recognition. Deep learning which is commonly used of performing voice recognition with good results because it
for speech recognition is a deep neural network (DNN). can determine the feature extraction in each layer.
Therefore, deep learning with the DNN model will be used
to detect speech recognition. The deep learning algorithm
used for this study uses several existing feature extractions.

IJISRT20OCT267 www.ijisrt.com 393

Volume 5, Issue 10, October – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
IV. METHODOLOGY 2) Spectral Bandwidth
Spectral Bandwidth of the wave which is the
The basic process for speech recognition systems maximum half of the crest represented by two vertical red
involves recording speech signals as input, preprocessing, lines and the wavelength axis. The fig shows. 3:
extracting certain features, and finally classifying them. This
research has 350 speech recording voice data.

A. Deep Neural Network

Deep Neural Network In three hidden layers is used
between the input layer and the output layer. The first
hidden layer has 256 units, the second hidden layer has 128
units, and the third hidden layer has 64 units. So that the
results of the summary of the model using the hard method
have the total number of parameters used in the deep neural
network is 48.194 or 48%. Shown in fig. 1.

Fig 3:-Bandwidth Spectral Features

3) Spectral Rollof
Spectral rollof to represent the frequency where the
high frequency drops to 0. A power spectrum where 85% of
the power is at the lower frequency. Fig4 shows:

Fig 1:- Input & output parameters of each layer

B. Dataset
This study uses a dataset[8] with male and female
genders. The total number of data consisted of 350 speakers,
each of whom had 190 men and 160 women who were
involved in this voice recording. The corresponding audio is
saved as a mono, 16bit, 32kHz WAV file. Fig 4:-Rollof Spectral Features
C. Extraction Feature 4) Zero-Crossing Rate
The voice recognition that we have previously Zero-Crossing Rate to measure the smoothness of the
processed is then extracted by several methods including signal is to count the number of zero-crossing in the signal
spectral centroid, spectral bandwidth, spectral rollof, MFCC segment. Sound signal oscillates slowly ega 100 Hz signal
(Mel Frequency Cepstral Coefficients), zero-crossing rate, will pass 100 zeros per second. Can be shown the fig. 5:
and feature chroma. But here we are focusing on feature
extraction for speech recognition using MFCC.

1) Spectral Centroid
Spectral centroid shows the energy frequency
spectrum indicating where the center of mass for the sound
is located. Shown the fig. 2:

Fig 5:-Zero-Crossing Rate Feature

Fig 2:- Centroid Spectral Features

IJISRT20OCT267 www.ijisrt.com 394

Volume 5, Issue 10, October – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
5) Chroma Feature Whereas for the output of each frame after filter
Chroma feature is for calculating chromatogramsfrom processing with N is the number of samples per frame Y [n]
a waveform or spectrogram. Chroma feature is a strong is the output signal X [n] and the input signal W [n] is the
spectrum for the representation of music audio, voices, etc. nth efficiency of the Hamming window [10]. Here is the
Shown in fig. 6: equation:

𝑌[𝑛] = 𝑋[𝑛] ∗ 𝑊[𝑛] (2)

Fast Fourier Transform (FFT) is used to convert the

signal from time to frequency[11].

The filterbankis based on the Mel scale, namely from

linear and logarithmic [12]. The following formula is used
to calculate Mel frequency:
𝑓
𝑀𝑒𝑙 (𝑓) = 2595 ∗ log 10(1 + 700) (3)

Fig 6:- Chromatogram or Spectrogram Feature The input from the filterbank of 2595 and 700 is a
fixed, predefined value that is widely used in the MFCC
6) MFCC method [13]. [14] and the last process is Discrete Cosine
Mel Frequency Cepstral Coefficients (MFCC) is one Transform (DCT) whose output is called Mel Frequency
method that is widely used in the field of speech technology, Cepstral Coefficients (MFCC). Shown in the fig.8:
both speech recognition, and voice recognition.

The Mel Frequency Cepstral Coefficient (MFCC)

technique is often used for the extraction of important
features from sound files based on different bandwidth
frequency for human hearing, the sound signal is filtered
linearly at low frequencies (below 1000Hz) and
logarithmically for high frequency (above 1000Hz).

Fig 8:-MFCC Feature

V. EXPERIMENT AND RESULT

The software instrument used to run the system is

cloud tools (Google Colab). The reason for using these tools
is due to the limited hardware used to support the system
processing speed. This section describes the deep neural
Fig 7:-MFCC Diagram Process network for speech recognition classification using SVM
and the overall results of speech recognition detection from
The first diagram process in MFCC is preemphasis, several matrices in terms of accuracy, precision, recall, and
which is to produce energy in high frequency which was f1. As explained in the following equation:
previously compressed during the process of producing
sound. Framing is used as trimming the sound signal file 𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝐹𝑃+𝐹𝑁+𝑇𝑁
(4)
which is divided into smaller parts, therefore signal analysis
can be processed by performing a short time (frame) in the 𝑇𝑃
recognition system. So it is most important to cut a signal 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (5)
𝑇𝑃+𝐹𝑃
that is smaller and still contains the original characteristics
for the signal analysis process [9]. Windowing is used to 𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = (6)
avoid interrupting signals that have been previously 𝑇𝑃+𝐹𝑁

processed. The function of the hamming window can be 2∗(𝑅𝑒𝑐𝑎𝑙𝑙 𝑥 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)

expressed in the equation: 𝐹1 = (𝑅𝑒𝑐𝑎𝑙𝑙+𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
(7)

2𝜋𝑛
𝑊[𝑛] = 0.54 − 0.46 cos[ ] (1)
𝑁−1

IJISRT20OCT267 www.ijisrt.com 395

Volume 5, Issue 10, October – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
In the above equation, it is TP true positive, TN true
negative, FP false positive, and FN false negative.

Some of these metrics will calculate the success of a

prediction whose accuracy results are 0.97%, recall 0.95%,
precision 0.88%, and F1 0.96%. Apart from several metrics,
a confusion matrix is also presented in this section. The
confusion matrix summarizes the classification results to
show the performance of the deep neural networks.

A. Support Vector Machine Classification

A support vector machine (SVM) is a conventional
learning model for pattern recognition and data analysis.
The SVM classification consists of a set of hyperplanes that
can be used for classification or regression analysis. After Fig 10:-Loss Classification Graph
the extraction feature is completewe use the SVM file to
measure the accuracy against the voice recognition This training is carried out after the feature extraction
classification. The results obtained from the SVM process is tested for its performance stability. This
classification, the accuracy is 0.9214 or 0.9%. SVM acts as performance stability is measured by the split percentage
one good approach to data modeling. Also, kernel mapping technique from 60%: 40% to 90%: 10%.
on vector machines provides a general basis as described in
table 1. The highest accuracy that occurs in the RBF kernel C. Confusion Matrix
is 0.9714% or 0.97%. This confusion matrix is a representative diagram that
is presented with a good process. This confusion matrix
Kernel SVM summarizes the complete classification results based on true
RBF Polynomial Linear and false objects. The correct classification result is 68 data.
Based on Figure 11. The label classification is marked with
0.9785 0.9642 0.9857 Female (0) and Male (1).
Insample accuracy

0.9714 0.9285 0.9714

Outsample accuracy
Table 1:- RBF,Poly & Linear Kernel Mapping

B. Training and Test Result

In this case, the training was conducted with 80% data
sharing for training data and 20% for testing data. As shown
in Figs. 9 and 10, is a graph of training accuracy, training
loss, validation accuracy, and validation loss. Each value is
training accuracy 0.7643, training loss 0.5276, validation
accuracy 0.7714, and validation loss 0.5643.

Fig 11:-Confusion Matrix

VI. CONCLUSION

In this study, the detection of voice recognition based

on gender used the DNN method which resulted in good
accuracy. Also, each speech recognition feature extraction
and classification. This extraction feature is to remove noise
in gender voice data by using MFCCbecause it is a good,
fast, and complete method. Meanwhile, the classification of
Fig 9:-Accuracy Classification Graph gender speech recognition uses the SVM algorithm which
has good accuracy results.

IJISRT20OCT267 www.ijisrt.com 396

Volume 5, Issue 10, October – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES

[1] Brigham E, The Fast Fourier Transform and Its

Application. Prentice-Hall Inc. New Jersey., 1988. .
[2] Y. Zhang, M. Pezeshki, P. Brakel, S. Zhang, C. Laurent, Y.
Bengio, and A. Courville, “Towards End-to-End Speech
Recognition with Deep Convolutional Neural Networks,”
In Proceedings of Interspeech, 2016. pp. 410-414.
[3] Jamil Ahmad, Mustansar Fiaz, Soon-il Kwon, Maleerat
Sodanil, Bay Vo, and Sung Wook Baik. Gender
Identification using MFCC for Telephone Applications - a
Comparative Study. arXiv preprint arXiv:1601.01577,
2016..
[4] Martin, A., and Przybocki, M. Speaker recognition in a
multi-speaker environment.In Proc. 7th European
Conference on Speech Communication and Technology
(Eurospeech 2001) (Aalborg,Denmark, 2001), pp. 787–90. .
[5] Tomi Kinnunen “Spectral Feature for Automatic Voice-
independent Speaker Recognition ”Depertment of
Computer Science, Joensuu University, Finland. December
21, 2003..
[6] L. Yufang, S. Jie, H. Wenjun, and P. Wenlin, "Speech
recognition of isolated word speech of primi language on
HTK," Journal of Yunnan Minzu University (Natural
Sciences Edition), pp. 426-430, 2015..
[7] M.-W. Lee and K.-C. Kwak, “Performance comparison of
gender and age group recognition for human-robot
interaction,” IJACSA) International Journal of Advanced
Computer Science and Applications, vol. 3, no. 12, 2012 .
[8] Akshay Babbar, "Akshay Babbar :
speakerrecognition,"2019. [Online]. Available:
https://www.kaggle.com/akshay4/speakerrecognition.
[Accessed: 19-Agustus-2020]., [Online].
[9] W. Junqin and Y. Junjun, “An improved arithmetic of
MFCC in speech recognition system,” in Electronics,
Communications and Control (ICECC), 2011 International
Conference on, 2011, pp. 719–722. .
[10] B. J. Mohan, “Speech recognition using MFCC and DTW,”
in Advances in Electrical Engineering (ICAEE), 2014
International Conference on, 2014, pp. 1–4. .
[11] A. Vijayan, B. M. Mathai, K. Valsalan, R. R. Johnson, L.
R. Mathew, and K. Gopakumar, “Throat microphone
speech recognition using mfcc,” in Networks & Advances
in Computational Technologies (NetACT), 2017
International Conference on, 2017, pp. 392–395..
[12] N. Alcaraz Meseguer, “Speech analysis for automatic
speech recognition,” Institutt for elektronikk og
telekommunikasjon, 2009. .
[13] S. T. Saste and S. M. Jagdale, “Emotion recognition from
speech using MFCC and DWT for security system,” in
Electronics, Communication and Aerospace Technology
(ICECA), 2017 International conference of, 2017, vol. 1,
pp. 701–704..
[14] G. Jhawar, P. Nagraj, and P. Mahalakshmi, “Speech
disorder recognition using MFCC,” in Communication and
Signal Processing (ICCSP), 2016 International Conference
on, 2016, pp. 0246–0250.

IJISRT20OCT267 www.ijisrt.com 397