Quran Recitation Recognition using End-to-End Deep Learning

Ahmad Al Harere*, Khloud Al Jallad

Department of Information and Communication Engineering,
Arab International University, Daraa, Syria.
*Corresponding author. E-mail(s): [email protected] Contributing authors: [email protected]

Abstract
The Quran is the holy scripture of Islam, and its recitation is an important aspect of the religion. Recognizing
the recitation of the Holy Quran automatically is a challenging task due to its unique rules that are not applied
in normal speaking speeches. A lot of research has been done in this domain, but previous works have detected
recitation errors as a classification task or used traditional automatic speech recognition (ASR). In this paper,
we proposed a novel end-to-end deep learning model for recognizing the recitation of the Holy Quran. The
proposed model is a CNN-Bidirectional GRU encoder that uses CTC as an objective function, and a character-
based decoder which is a beam search decoder. Moreover, all previous works were done on small private
datasets consisting of short verses and a few chapters of the Holy Quran. As a result of using private datasets,
no comparisons were done. To overcome this issue, we used a public dataset that has recently been published
(Ar-DAD) and contains about 37 chapters that were recited by 30 reciters, with different recitation speeds and
different types of pronunciation rules. The proposed model performance was evaluated using the most
common evaluation metrics in speech recognition, word error rate (WER), and character error rate (CER). The
results were 8.34% WER and 2.42% CER. We hope this research will be a baseline for comparisons with future
research on this public new dataset (Ar-DAD).

Keywords: Deep Learning, End-to-End, Speech Recognition, Quran recitation, Natural Language Processing

1. Introduction
Speech communication is an important way of social interaction to convey our thoughts, ideas, and emotions
to others. Moreover, speech is also a crucial tool for learning and education, as it is the primary way in which
information is exchanged between teachers and students.

Processing speech using computers and artificial intelligence is a complex task that has been a hot research
topic in recent years. One of the main challenges is automatically transcribing spoken words into text. The
performance of such systems has greatly improved in recent years due to advancements in deep learning.

Arabic language is a rich language with a long history and cultural significance. It is spoken by over 400 million
people worldwide, and it is the official language in many countries. Arabic language recognition is a challenging
task because of a lack of resources and, having many variations in pronunciation and dialects. Although these
difficulties, researchers have made significant progress in developing Arabic speech recognition systems,
which can be used in tasks such as automatic speech transcription and translation, and also in speech-enabled
applications such as voice assistants and chatbots.

The Holy Quran, which was revealed in Arabic language, holds a central place in the hearts and minds of
Muslims. Quran is considered the holy book of Islam and the words of Allah. In addition, it is a guide for all
aspects of life, providing moral and spiritual teachings and a source of inspiration and guidance. The Holy
Quran recitation recognition is a particularly challenging task because of the specific requirements of the task,
such as recognizing different recitation styles and checking the correct pronunciation of Tajweed rules, a set of
pronunciation rules that must be applied to recite the Quran in the same way that the Prophet Muhammad did.
Also, Quran recitation includes many unique sounds and intonations that are not used in other forms of Arabic

1
spoken language. Many researches have been done on Quran recitation processing over time to make Quran
recitation easier and more accessible to a wider audience. One of these hot topics of researches is Quran
recitation recognition systems for tasks such as recognizing reciters and detecting errors in recitation based
on Tajweed rules.

Although the research topic of recognizing the recitation of the Holy Quran has been a hot topic in recent years,
most of research papers are limited to detecting mispronunciation of words or some Tajweed rules on private
small datasets. Some researchers proposed detecting the mispronunciation of reciting some verses directly
from speech features, while others proposed converting the recitation into text using traditional ASR. As all
previous works were conducted on small private data sets of few chapters, they do not cover a large number
of examples of Tajweed rules and different speeds of recitation. As a result, previous works are still not effective
to be applied in real-life applications because detecting only the mispronunciation of the Tajweed rules is less
important than detecting the mispronunciation of words. Moreover, the use of traditional ASR in recitation
recognition suffers from many problems as ASR models require specific forms of datasets which are not
available for Quran recitations yet. We will discuss those problems in detail later in the related works section.

This paper aims to fill these gaps through the use of end-to-end deep learning methodologies that overcome
the problems of using traditional ASR. Experiments were done on the Ar-DAD dataset, which is a large public
dataset covers most of Tajweed rules and the different speeds of recitation through the participation of about
30 readers in reciting the verses.

The main contributions of this work are:

• Using the end-to-end methodology instead of the traditional ASR, for automatic phoneme alignment,
so we do not need alignment tools anymore. To the best of our knowledge, the task of Quran recitation
recognition has not yet been tackled using an end-to-end deep learning approach, and this work is to
fill this gap.
• We have evaluated our model on a big public dataset, so it can be a baseline model for comparisons
later.
• By comparing the predicted text with the real text, our solution can determine the type of error
(deleting, substituting, adding) at the level of words, characters, and some Tajweed rules, with the
exact position of the error.

In terms of limitations, first, all the samples in the Ar-DAD dataset are for men reciters, which makes the model
less robust for recognizing the recitations by women and children. Second, the Ar-DAD dataset contains
samples from only one recitation form, the 'Hafs from Aasim' recitation [1][2]. However, recitation of the Holy
Quran has ten recitation forms (Qira'at) approved by scholars [3]. The differences between these ten forms are
mainly in the pronunciation of certain words, prolongation, and intonation [1]. Each form of recitation has its
own unique features that distinguish it from the others. This may lead the model to incorrectly recognize
samples with different forms of recitation.

2. Related Works
The Arabic language has several forms, two formal forms, and many slang forms. As for formal forms, Arabic
has the classical Arabic language (CA), which is the language of the Holy Quran, and the modern standard Arabic
(MSA) that is used in, news, books, … etc. As for the slang forms, Arabic has many dialects that differ from one
country to another. Since this paper is for the recognition of classical Arabic speech, the literature review will
focus only on papers done for the recognition of the Holy Quran recitation. But there are no researches done
using end-to-end on Holy Quran, so we discussed some researches that used end-to-end deep learning
methodologies on Modern Standard Arabic.

2
There are several challenges in Quran Recitation Recognition. First, there are several letters In the Arabic
language that have confusing pronunciation, as they share the articulation way out and some
characteristics.[4], [5]. This confusion significantly impacts speech recognition models, as it leads to errors in
the recognition of Arabic speech and decreases the overall accuracy of the model. For example, suppose a model
is not able to distinguish between the letters " ‫( "ص‬ṣād) and "‫( "س‬seen). In that case, it may incorrectly
recognize the word "‫( "الصراط‬alserat) as "‫( "السراط‬alserat) which can lead to misinterpretation of the text and
an increase in both Word and Character Error Rates. Second, the complex and nuanced rules of Tajweed make
it difficult for an AI model to accurately recognize recitation because some of these rules change the
pronunciation way of letters when applied. For example, the Turning rule ( ‫ )اإلقالب‬when applied, the
pronunciation of "‫( "ن‬noon) becomes "‫( "م‬meem). Third, having different forms of Quran Recitation make the
ASR more complex task because of different ways of pronouncing some of the Tajweed rules, and the length or
manner of pronunciation may vary depending on the context and the reciter. For example, Separated
Lengthening (‫)المد المنفصل‬, which is the prolongation of a letter that comes at the end of a word, can be
pronounced for 2, 4, or 5 counts in length. Another example is Concealment (‫)اإلخفاء‬, which is the rule of hiding
the pronunciation of certain letters, it can be pronounced in different degrees of hiddenness.

Moreover, one of the main difficulties in the Quran recitation detection task is that there are three different
speeds for reciting the Quran: Hadr (‫)حدر‬, Tahqeeq (‫)تحقيق‬, and Tadweer (‫[ )تدوير‬6]. Each speed has its own
unique advantages, and each of them is used to help listeners understand the Quran better and to get the most
out of the recitation.

Hadr (‫)حدر‬: is typically considered the fastest speed of recitation, where the emphasis is on fluency and the
ability to recite large portions of the Quran quickly and smoothly. This speed is particularly useful for those
who are already familiar with the Quran, have a deep understanding of the Quran, and are proficient in the
rules of Tajweed.

Tadweer (‫)تدوير‬: is the moderate speed of recitation, where the emphasis is on proper pronunciation and
intonation, while still maintaining a relatively moderate pace. This speed is particularly useful for those who
have a basic knowledge of Tajweed and are trying to improve their recitation and pronunciation.

Tahqeeq (‫)تحقيق‬: is the slowest speed of recitation, where each letter is pronounced clearly and deliberately,
allowing the listener to fully understand the meaning of each verse. This speed is particularly useful for those
who are learning to recite the Quran for the first time or for those who are still not familiar with the rules of
Tajweed.

In this section, we will discuss two basic types of researches, researches on Quran and researches that used
end-to-end deep learning on Modern Standard Arabic (MSA) as there are no end-to-end experiments on Quran.

As for researches on Quran, we have two basic types of methodologies.

• Researches based on detecting mispronunciation from speech directly. Either for Tajweed Rules
Mispronunciation or Character Mispronunciation. Table 1 shows a comparison of these studies.
• Researches based on traditional ASR that convert speech to text, then detect mispronunciation by
comparing the result text with Quran text.

2.1 Researches based on detecting mispronunciation from speech directly

Researches based on detecting mispronunciation from speech directly were done using traditional
methodologies, such as Hidden Markov Model (HMM) and Gaussian Mixture Model (GMM), or machine learning
models such as Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP), as follows:

3
Hassan et al. [7] developed a solution to recognize Qalqalah Kubra1 (‫ )القلقلة الكبرى‬pronunciation using Multilayer
Perceptron as a classifier and MFCC as features extraction. The dataset used contains 50 samples, each with
correct and incorrect pronunciation, and the achieved results ranged from 95% to 100%.

Al-Ayyoub et al. [8] used machine learning to build a model for the automatic recognition of Quran Recitation
Rules (Tajweed). This model was able to determine the recitation correctness of the following eight rules of
intonation: EdgamMeem2, EkhfaaMeem3, Tafkheem Lam4, Tarqeeq Lam5, Edgam Noon6(Noon), Edgam Noon
(Meem), Edgam Noon (Waw) and Edgam Noon (Ya’). The authors used a dataset that consists of 3,071 audio
files, each containing a recording of exactly one of the eight rules under consideration (in either the correct or
the incorrect usage of the rule). For feature extraction, many feature extraction techniques were used such as
Linear Predictive Code (LPC), Mel-frequency, Cepstral Coefficients (MFCC), Multi-Signal Wavelet Packet
Decomposition (WPD), and Convolutional Restricted Boltzmann Machines (CRBM). As for classification, several
classifiers were used such as k-Nearest Neighbors (KNN), Support Vector Machines (SVM), Artificial Neural
Networks (ANN), and Random Forest (RF), with accuracy of 96% using SVM.

Alagrami et al. in [9] proposed a solution that makes use of threshold scoring and support vector machine
(SVM) to automatically recognize four different Tajweed rules (Edgham Meem, Ekhfaa Meem, takhfeef Lam,
Tarqeeq Lam) with 99% accuracy, where the filter banks were adopted as feature extraction. The dataset used
contained about 657 records of Arabic natives and non-natives, each rule has 160 records, and each of them is
either the correct pronunciation or the wrong pronunciation of this rule.

A Tajweed classification model was developed by Ahmad et al. in [10]. This solution focused on a set of Tajweed
rules called “the Noon Sakinah rules” and in particular the rule of “Idgham” with and without “Ghunnah”. Mel-
Frequency Cepstral Coefficient and a neural network were used for the feature extraction and the classification
process, where Gradient Descent with Momentum, Resilient Backpropagation, and the Levenberg-Marquardt
optimization algorithms, were used to train the neural network. The Levernberg Marquardt algorithm achieved
the highest test accuracy (77.7%), followed by Gradient Descent with Momentum (76.7%) and Resilient
Backpropagation (73.3%). The dataset used was 300 audio files of recitation of two famous reciters, and each
is a recitation of one of those Tajweed rules.

Nahar et al. in [11] took a different path as they proposed a recognition model to recognize the “Qira’ah” from
the Holy Quran recitation precisely 96%, since according to the narration "hadith" No. 5041, taken from [12],
the Holy Quran has seven main reading modes, known as "Qira'at," which are acknowledged as the most
popular methods of reciting from the Holy Quran, and three complementary readings of the seven. This model
used the Mel-Frequency Cepstrum Coefficients (MFCC) features and Support Vector Machine (SVM), where the
authors have built a dataset has 10 categories, each one representing a type of Holy Quran recitation or
“Qira’ah”, with a total of 258 wave files.

For detecting letter and word errors from speech directly, the researches carried out in this field used
classifiers trained on datasets containing samples of mispronunciation and correct pronunciation of specific
verses, or they stored the characteristics of correct recitation of certain verses in a database. Compare users'
recitations with stored recitations and then calculate the similarity using a threshold. One of the most

1 Qalqalah: it is the vibration of sound at the end of the pronunciation of a letter.

The letters of Qalqalah: ‫قطب جد‬.
2 EdgamMeem :If the start of the word begins with a meem and is followed with a meem sakinah, then merge the words through

the meem and apply ghunnah.

3 EkhfaaMeem: If a [‫ ]ب‬is followed after a meem sakinah, then apply a ghunnah while hiding the meem sakin before continuing to the [‫]ب‬.
4 Tafkheem Lam: If there is a Fatha or a Dhamma before the word of Allah or Allahum, then laam in Allah will be heavy.
5Tarqeeq Lam: If there is a kasrah ( ِ) before the word Allah, then the Laam in Allah or Allahum will be light.
6Edgam Noon: If the Noon Sakin ( ِ) or the Tanween is followed by any of Ingham letters, the reciter should skip the Noon or the

Tanween and pronounce the following letter.

4
significant shortcomings of this methodology is that it can only categorize the recitations of the verses that
were presented in the datasets.

One of the earliest works done in this field is the work of Tabbal et al. in [13], where an automated verse
delimiter and an error detection model were developed for the recitation of the Holy Quran. HMM classifier
and Mel Frequency Cepstral Coefficient (MFCC) features were used on a private dataset that is one hour of
recitations of Surah Al-Ikhlas only. The best accuracy obtained by this solution was 85% for females and 90%
for males.

Putra et al. [14] developed software for Quranic Verse Recitation Learning. The solution proposed used Mel
Frequency Cepstral Coefficient (MFCC) for feature extraction and the GMM model as a classifier. In order to test
the reliability and accuracy of correction, a data set was collected from ten speakers reading some verses
incorrectly and correctly for each of them. The achieved correction accuracy was 90% for hija’iyah letters
(Arabic Alphabet Letters) pronunciation, 70% for recitation law where the law might be idgham, ikhfa’ or idhar,
and 60% for the combination of pronunciation and recitation law.

In [15], Rahman et al. proposed an automated checking system to learn children the correct recitation of the
Holy Quran. Mel-Frequency Cepstral Coefficient (MFCC) was used for feature extraction, and Hidden Markov
Model (HMM) was used for classification and recognition. Using the HMM algorithm, the model can identify
and highlight any discrepancy or inconsistency in children's recitation by comparing it with the correct
teacher's recitation that was stored in a database, where only one chapter of the Quran was supported, Surah
Al-Fatiha.

Muhammad et al. [16] proposed E-Hafiz system to facilitate reciting learning of the holy Quran, where Mel
Frequency Cepstral Coefficient (MFCC), Vector Quantization (VQ), and Calculation of distance between vectors
were used to extract the features, reduce the number of features vectors and compare the result with the
threshold value. A dataset of 10 expert recitations of the first 5 surahs of the Holy Quran was used, and
recognition accuracy of 92%, 90%, and 86% was achieved for men, children, and women, respectively.

Rajagede and Hastuti [17] propose a model to help users to check their Al-Quran memorization using the
Siamese Long Short-Term Memory (LSTM) Network. Siamese LSTM network was used to check the similarity
between two samples, so it verifies the recitation by matching the input with existing data for a verse read,
without performing a speech-to-text extraction process. Two Siamese LSTM architectures were compared: the
Siamese-Classifier, which employed binary classification, and the Manhattan LSTM, which produced a single
numerical value to indicate similarity. In addition, the performance of the models was compared with Mel
Frequency Cepstral Coefficient (MFCC), Mel Frequency Spectrum Coefficient (MFSC), and Delta Features, where
an F1-score of 77.35% was given by using MFCC with delta features and Manhattan LSTM, as the best result
obtained. Four reciters who recited 48 verses from the last 10 Surahs of the Quran provided the data used to
train the model.

2.2 Researches based on traditional ASR

Very few researches suggested converting the recitation speech into text by using Automatic Speech
Recognition techniques, then detecting mispronunciation by comparing the result text with Quran text. This
methodology is better than models that detect mispronunciation from speech directly, as it helps to detect
various errors in the user's recitation by comparing the predicted text with the original text of the verses.
Moreover, no need for wrong pronunciation samples.

To detect mistakes made during the recitation of the Holy Quran, Tabbaa and Soudan [18] created a computer-
aided recitation training solution that combined Automatic Speech Recognition (ASR) with a classifier-based
approach to increase the detection rate. This solution detects errors in two phases: the HMM-based ASR
recognizes the recitation, and then the classifier applies, where two classifiers were used, one to distinguish
between the stressed and non-emphasized pronunciations of the Arabic letter "R", and the other to separate

5
closely related and frequently mixed-up letter pronunciations. The HMM recognizer was trained using CMU
Sphinx, and the classifiers were built using WEKA (Waikato Environment for Knowledge Analysis), where
numerous machine learning algorithms were tested. Up to 7 hours of recitations were recorded by phone calls
to a TV program, the recitation scholar reads a page from the Quran before listening to the students’ recitations
and correcting any mistakes. According to the results, the system has a word-level accuracy of 91.2%, where it
has been tested on 60 minutes of continuous recitation.

Al-Bakeri et al. in [19] introduced an ASR integrated with a self-learning environment that depends on MVC
architectures to correct recitation automatically. The speech recognition model was built using open-source
CMU Sphinx tools, which also contain the Hidden Markov Model (HMM) code that was chosen for feature
extraction, feature training, and pattern recognition. Also, language models were used in the process of building
the system and built using CMU-CSLMT tools [20]. The corpus contains the recitations of two short chapters,
Surah Al-Ikhlas and Surah Al-Rahman, which were recorded by 10 famous Quran reciters. To assess the ASR
performance, the word error rate (WER) was used, where the ASR output was compared with the correct words
of the verse considering insertions, deletions, and substitution, so the correctness was presented as a number
ranged between 47.47% and 75.2% as reported.

A speech recognizer for the Holy Quran was introduced by Tantawi et al. in [21]. This solution is able to
recognize the recitation of some verses in addition to some Tajweed rules that were taken into account during
the development process, where it was trained using 32 recordings of Chapter 20 from the Holy Quran
according to Narration of “Hafs on the authority of Asim” (One of the ten reading forms of the Quran). The
pronunciation dictionary for the Holy Quran recitations was built using an automated tool proposed by [22],
where the transcription was passed to it to build the dictionary. As for the language model, The SRI Language
Modeling (SRILM) toolkit [23] was used. With the KALDI toolkit, numerous experimental configurations with
various dataset sizes and Tajweed rules were used. The best experimental setup used MFCC features and Time
Delay Neural Networks (TDNN), where Word Error Rates (WER) and Sentence Error Rates (SER) ranged from
0.27 to 6.31% and 0.4 to 17.39%, respectively.

However, this methodology is not effective in recognizing Quran recitation as it is based on traditional ASR,
because there was a problem with the alignment process needed to train the acoustic model. The traditional
ASR consists mainly of three models:

• Pronunciation dictionary that converts words from the original language into a series of phonemes
that express the pronunciation of these words.
• Acoustic model that connects phonemes with the features extracted from the corresponding sound.
• Language model that is responsible for determining the most likely sequence of words based on the
context and grammar of the language.

Figure 1 shows how these components together:

6
 Fig. 1 Traditional ASR workflow [24]

Training datasets must contain the correct alignment between the acoustic frames and phonemes in order to
train the acoustic models. This is one of the biggest problems in this field, as no dataset of this format is available
for Quran recitations, unlike the MSA, which has several datasets for this format, such as the KAPD dataset [25]
and the Nawar Halabi dataset [26]. As a result, researchers who proposed this method used automated tools
that perform this alignment process, but the results were not good enough to recognize the recitation
efficiently.

Fig. 2 Traditional ASR pipeline [27]

Fig. 3 End-to-End ASR pipeline [27]

For this reason, we proposed using the end-to-end methodology instead of the traditional ASR. The end-to-end
models can do the alignment process automatically without any need for additional tools, and convert acoustic
7
features to text transcription directly without the need for all other components needed in traditional ASR,
which makes them more efficient and suitable for Quran recitation recognition. Figure 2 and Figure 3 show a
comparison between the conventional ASR pipeline and the end-to-end pipeline. To the best of our knowledge,
the task of Quran recitation recognition has not yet been tackled using an end-to-end deep learning approach,
and this work is to fill this gap.

Table 1 A comparison of some recitation recognition works

Work Dataset Feature Model Recognition Level Accuracy

Extraction
[7] 100 samples for Mel- Multi-Layer One Tajweed rule 95% - 100%
Qalqalah Frequency Perceptron (Qalqalah)
pronunciation Cepstral (MLP)
Coefficient neural
(MFCC) network

[8] 3071 audio CRBM, Support Eight Tajweed rules 96.4%

files, each MFCC, Vector (EdgamMeem,EkhfaaMeem,
containing a WPD, Machine Tafkheem Lam, Tarqeeq
recording of HMM-SPL (SVM) Lam, Edgam Noon (Noon),
exactly one of Edgam Noon(Meem),
the eight rules Edgam Noon(Waw) and
Edgam Noon(Ya’))

[9] 657 records of Filter Banks Support Four different Tajweed 99%
Arabic natives Vector rules (Edgham Meem,
and non-natives Machine Ekhfaa Meem, takhfeef
(SVM) Lam, Tarqeeq Lam)

[10] 300 audio files Mel- Neural “Idgham” rules with and 73.3% - 77.7%
of recitation of Frequency Network without “Ghunnah”
two famous Cepstral
reciters Coefficient
(MFCC)

[11] 258 wave files Mel- Support Classification of recitation 96%

Frequency Vector into one of ten recitations
Cepstral Machine types (“Qira'at”)
Coefficient (SVM)
(MFCC)

[14] Voice recorded Mel- Gaussian Letter level and some Letters: 90%
from an expert Frequency Mixture Tajweed rules Tajweed rules: 70%
Cepstral Model Combination: 60%
Coefficient (GMM)
(MFCC)

[16] Consists of 10 Mel- Threshold Word level for verses in the 86% - 92%
expert Frequency based on dataset
recitations of Cepstral Euclidean
the first 5 Coefficient distance
surahs of the (MFCC)
Holy Quran.

8
 [13] About 1 hour of Mel- Hidden Sourate Al-Ikhlass with the 85% - 90%
audio Frequency Markov most important
recitations of Cepstral Model Tajweed rules
Sourate Al- Coefficient (HMM)
Ikhlass (MFCC) based on
Sphinx

[28] Some specific Mel- Threshold Word level for verses in the 90% - 92%
verses recited Frequency based on dataset
by experts Cepstral distance
Coefficient
(MFCC)

[17] 48 verses from Mel- Siamese 48 verses from the last 10 77.35%
the last 10 Frequency LSTM Surahs
Surahs recited Cepstral
by four reciters Coefficient
(MFCC)

As our proposed model is based on an end-to-end model and there is no such model for Quran recitation
processing, we will discuss it in the MSA since there are some researchers who used end-to-end deep learning
on modern standard Arabic. A comparison of these works is shown in Table 2.

Hussein et al. [29] proposed an End-to-End transformer-based Arabic Automatic Speech Recognition (ASR)
model with a multitask objective function of Connectionist Temporal Classification (CTC)/Attention, where
long short-term memory (LSTM) and transformer-based language model (TLM) were the two kinds of language
models utilized in this work. The proposed model was compared to the previous approaches for Modern
Standard Arabic (MSA) recognition task using Multi-Genre Broadcast 2 (MGB2) [30] data and for the Dialectal
Arabic recognition task using MGB3 [31] and MGB5 [32] data. While the conventional word error rate (WER)
was used to evaluate the model results for the first task, the multi-reference word error rate (MR-WER) and
averaged WER (AV-WER) which adopted from MGB3 [31] and MGB5 [32] challenges, were used to evaluate the
model results for the second task. 12.5%, 27.5%, 33.8% WER were achieved for the MGB2 [30], MGB3 [31], and
MGB5 [32] challenges, respectively.

Ahmed et al. [33] introduced an end-to-end model based on Bidirectional Recurrent Neural Network with CTC
objective function and a 15-gram language model as an Arabic speech-to-text transcription system. Also, a
character‐based decoder without a lexicon was used. This model was evaluated using 1200 hours corpus of the
Aljazeera multi-Genre broadcast programs (MGP2) [30], where the WER was 12.03% for non-overlapping
speech on the development set.

Belinkov et al. [34] analyzed the internal learned representations in an end‐to‐end ASR model for two
languages (English and Arabic). Three datasets were used, Librispeech [35] and TED-LIUM [36] were used for
English and the MGB‐2 corpus [30] which has 1200 h from the Al Jazeera Arabic TV channel, was used for
Arabic.

Alsayadi et al. [37] proposed end‐to‐end deep learning approaches to build a diacritized Arabic ASR. Two types
of speech recognition approaches were used: The conventional ASR approach and the end‐to‐end ASR
approach which consists of two models. The first model was built using Joint CTC attention based on the ESPnet
toolkit [38] with an RNN-based language model, and the second model was built based on CNN‐LSTM with the
attention method using the Espresso toolkit [39] and with an external LM containing about 1.8 m words and
245k unique words. Training and testing of these models were done based on the Standard Arabic Single
Speaker Corpus (SASSC), which contains 7 h of modern standard Arabic speech. The WER of 33.72%, 31.10%,
9
and 28.48% were achieved for conventional ASR, the first end-to-end model, and the second end-to-end model,
respectively.

Table 2 Comparison between works related to end-to-end approaches for Arabic ASR

Work Dataset Model Language Model WER

[29] MGB2, MGB3, transformer-based with a LSTM and transformer- MGB2: 12.5%
and MGB5 multitask objective function based MGB3: 27.5%
of CTC/Attention MGB5: 33.8%

[33] MGP2 Bidirectional Recurrent 15-gram 12.03%

Neural Network based with
CTC objective function

[37] SASSC Conventional ASR 3‐gram 33.72%,

Joint CTC attention RNN based 31.10%

CNN‐LSTM with attention external LM 28.48%

3. Dataset
The dataset used in this work is the Ar-DAD dataset [40] which is a large dataset of Arabic-based audio clips
containing 15810 clips of 30 popular reciters reading 37 chapters from the Holy Quran in addition to 397 audio
clips for 12 imitators of the top reciters and two plain text files that contain the same chapters' textual content
read by the reciters with and without vocalization (vowelization).

The audio samples, which are 10 seconds long on average and have a sampling rate of 44.1 kHz, 16-bit depth,
and stereo channels, are shared in the WAV format.

The dataset was split as 80% for training, 10% for testing, and 10% for validation, where 12648 clips, 1581
clips and 1581 clips, were selected randomly and used for training, testing, and validation, respectively.

We noticed that the dataset contains all speeds of recitations mentioned before. In addition, the majority of the
first verse samples transcripts of each chapter contain the sentence “‫الرحِ ِيم‬ َّ ‫ ” ِبس ِْم اللَّـ ِه‬while the
َّ ‫الرحْ َم ٰـ ِن‬
corresponding audio clips do not contain the pronunciation of this sentence, so we removed this sentence from
all transcripts because it would cause a problem in training the model since the number of these samples is
about 1100 samples out of 15810 samples.

4. Methodology
The proposed solution consists of two main components: a CNN-Bidirectional GRU encoder and a character-
based decoder. The encoder maps the input vector of features to a latent representation. The decoder takes the
latent representation and generates one prediction at a time. CTC is the objective function used to train the
encoder. The next subsections discuss them in detail.

4.1 Encoder
The encoder is CNN-BiGRU. The reason behind using CNN as the first layer is that the ASR performance can be
improved by applying convolutions in frequency and time domains to spectral input features [41]–[43]. In
addition, using Bidirectional RNNs in speech recognition provides better context utilization, both forward and
backward, to accurately predict words [44].

10
The input of the encoder is the normalized spectrogram of audio clips. Each audio clip is a time series of length
𝑇 with a vector of audio features for each time slice. The input vectors 𝑉1 , 𝑉2 , . . . , 𝑉𝑇 , are prepared by the 2D
convolution layers (time and frequency domains) and then the CNN output is fed as input to Bidirectional GRUs.
The output probabilities of the encoder are maximized using the CTC loss function.

4.1.1 Convolution Neural Network (CNN)

Convolutional Neural Networks (CNNs) are a type of deep learning architecture widely used for image
classification and recognition tasks [45]. CNNs are designed to automatically learn the features of the input
data and make predictions based on those features. The input data is typically processed in a series of
convolutional layers, activation functions, and pooling layers.

In a convolutional layer, a set of kernels, also known as filters, slide over the input data, computing dot products
between the input data and the weights of the kernels. The dot products are then used to produce feature maps,
which are fed into the activation functions to introduce non-linearity to the model. The size of the kernels, the
stride (the step size at which they move over the input data), and the padding (the addition of zeros around the
input data to control the size of the output feature maps), are all hyperparameters that can be optimized for
the specific task and input data.

2D Convolutional Neural Networks (2D-CNNs) are a specific type of CNN that operate on 2D input data, such
as an image. In a 2D-CNN, the convolutional layers perform 2D convolutions, using 2D kernels to scan the input
data and extract local features. The hierarchical representations built by the multiple convolutional and pooling
layers allow 2D-CNNs to learn increasingly complex and discriminative features, making them well-suited for
tasks such as object recognition and segmentation[46]. In addition, 2D-CNNs can be a powerful tool for speech
recognition tasks, as they allow for the automatic extraction of relevant features from the spectrogram of an
audio signal. By combining 2D-CNNs with other deep learning architectures, such as RNNs, end-to-end speech
recognition systems can be created that can handle variable-length inputs and model the complex relationships
between speech sounds [47].

4.1.2 Bidirectional Recurrent Neural Networks (Bi-RNNs)

A Bi-RNN is a combination of two RNNs that process two sequences, one in a forward direction and one in a
backward direction, to benefit from the information of the past and the future and to compute the likelihood of
the output character 𝑐 at a given time input 𝑋𝑡 , depending on the previously hidden state ℎ𝑡−1 , the current input
𝑋𝑡 and the next hidden state ℎ𝑡+1 [44].

Fig. 4 Bidirectional Recurrent Neural Network [48]

As shown in Figure 4, there is an additional-hidden layer for each Bi-RNN layer to accommodate the backward
training process, where the forward and backward hidden states are updated at a given time 𝑡 as follows:
11
 𝐴𝑡 (𝐹𝑜𝑟𝑤𝑎𝑟𝑑) = 𝑓 (𝑋𝑡 ∗ 𝑊𝑋𝐴
𝑓𝑜𝑟𝑤𝑎𝑟𝑑
+ 𝐴𝑡−1 (𝐹𝑜𝑟𝑤𝑎𝑟𝑑)) ∗ 𝑊𝐴𝐴
𝑓𝑜𝑟𝑤𝑎𝑟𝑑 𝑓𝑜𝑟𝑤𝑎𝑟𝑑
+ 𝑏𝐴 (1)

𝑏𝑎𝑐𝑘𝑤𝑎𝑟𝑑
𝐴𝑡 (𝐵𝑎𝑐𝑘𝑤𝑎𝑟𝑑) = 𝑓 (𝑋𝑡 ∗ 𝑊𝑋𝐴 𝑏𝑎𝑐𝑘𝑤𝑜𝑟𝑑
+ 𝐴𝑡+1 (𝐵𝑎𝑐𝑘𝑤𝑎𝑟𝑑)) ∗ 𝑊𝐴𝐴 + 𝑏𝐴𝑏𝑎𝑐𝑘𝑤𝑎𝑟𝑑 (2)

Where 𝑏 is the bias, 𝑊 is the weight matrix and 𝑓 is the activation function. And the hidden state is:

ℎ𝑡 = 𝐴𝑡 (𝐹𝑜𝑟𝑤𝑎𝑟𝑑) + 𝐴𝑡 (𝐵𝑎𝑐𝑘𝑤𝑎𝑟𝑑) (3)

A Bi-RNN's network block can be vanilla RNNs that suffer from vanishing gradient [49] and exploding gradient
problems [50], Gated Recurrent Units (GRU), or Long Short-Term Memory (LSTM). GRU and LSTM
architectures are the most used when using long RNNs, because they handled vanishing and exploding gradient
problems and are capable of learning long-term dependencies [51], [52]. In this work, we used GRU
architecture because it has fewer parameters than LSTM and is faster to train (a GRU has two gates, reset and
update gates, whereas an LSTM has three gates, input, output and forget gates). Also, it was discovered that the
performance of GRU and LSTM were comparable for some tasks involving speech signal modeling and natural
language processing [53], [54].

Figure 5 shows the architecture of a GRU cell and the equations to calculate the value of gates.

Fig. 5 GRU architecture and equations [55]

Where 𝑧 and 𝑟 represent the update and reset gates respectively, 𝜎 is the sigmoid function, 𝑋𝑡 is the current
input, 𝑊 is the matrix of weights, ĥ is the current memory content and ℎ is the final memory at the current time
step.

The Bi-RNNs layers are followed by a fully connected layer and an output layer which uses the softmax function
to calculate the probability distribution over characters as follows:

exp(𝑤𝑘𝐿 . ℎ𝑡𝐿−1 ) (4)

𝑝(𝑐 = 𝑘|𝑥) =
∑𝑗 exp(𝑤𝑗𝐿 . ℎ𝑡𝐿−1 )

𝐿 represents the output layer, so ℎ𝐿−1 is the hidden representation of the previous layer.

4.1.3 Connectionist Temporal Classification (CTC)

Connectionist Temporal Classification (CTC) [56] is the loss function used to train the model. CTC is an output
and scoring function that addresses sequence problems when the alignment between the input and the output
is not known, so it is applied in applications like speech and handwriting recognition.

For each time step 𝑡 and a single input sequence 𝑋 of length 𝑇, the encoder gives a distribution over the
vocabulary, 𝑝𝑡 (𝑐|𝑋), then CTC computes the probability for a single sequence 𝐶 of length 𝑇 as follows:
12
 𝑇
(5)
𝑃(𝐶|𝑋) = ∏ 𝑝(𝑐𝑡 |𝑋)
𝑡=1

The same word can be represented by several different sequences, so finding the most likely sequence is done
by summing over the probability of these sequences:

𝑃(𝑆|𝑋) = ∑ 𝑃(𝐶|𝑋) (6)

Finally, CTC loss, which is the negative log probability of all valid sequences, is calculated using the dynamic
programming algorithm which speeds up the calculation. In addition, to calculate its derivative, which utilizes
the backpropagation through time algorithm to update the encoder's parameters. CTC loss function:

𝐿𝐶𝑇𝐶 = − log 𝑃(𝑆|𝑋) (7)

4.2 Decoder
In this paper, a character-level decoder was used because this type of decoder has several advantages over
word-level decoders. One of the main advantages is that character-level models are more robust to out-of-
vocabulary (OOV) words and variations in pronunciation [57]–[59]. Since the model is trained to predict
individual characters, it is able to handle words that it has not seen before by predicting the individual
characters that make up the word. This is particularly useful in speech recognition, where there may be many
rare or unknown words. Character-level decoders also tend to be more computationally efficient than word-
level decoders [60]. Since the model is only predicting individual characters, it does not need to search through
a large vocabulary to find the most likely word. This can make decoding faster and more efficient.

In general, decoders are used to find the proper output for a given input by solving the following equation:

𝑆 ∗ = 𝑎𝑟𝑔𝑚𝑎𝑥𝑆 𝑝(𝑆|𝑋) (8)

Greedy algorithms are used to solve this problem by using the output that is most likely at each time step.
However, these algorithms have a big problem, which is overlooking the possibility that a single output
sentence could have a variety of alignment forms [57]. For that, we used the CTC beam search decoder that
sums the probabilities of each sentence to produce the best result. The CTC loss function is used to train the
model, and the beam search decoder is used to generate the final output sequence. The beam search decoder
works by maintaining a fixed number of top-scoring sequences (the "beam") at each decoding step, rather than
considering all possible next steps. This reduces the search space and allows for faster decoding while still
maintaining good accuracy. Additionally, the CTC loss function allows the decoder to be robust to variations in
the timing of the input, making it well-suited for speech recognition tasks. Overall, the CTC beam search decoder
is an efficient and effective method for decoding sequences in speech recognition and other sequence-to-
sequence tasks [61].

5. Experiment Setup
Several speech recognition models that have shown highly accurate results were proposed in recent years.
One such model is Deep Speech 2 [62]. Deep Speech 2 was developed by Baidu Research, it uses a combination
of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to transcribe speech to text,
as shown in Figure 6. The model starts with a convolutional layer that extracts features from the spectrogram,
followed by several more convolutional layers that extract higher-level features. The output of these layers is
then passed through a stack of bidirectional Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU)
layers to process the sequential information in the speech, then a linear layer maps the output to the final text
transcript. The training uses a Connectionist Temporal Classification (CTC) loss function that allows the model
to align the output labels to the input speech, regardless of the length mismatch between the two.

13
 Fig. 6 Architecture of Deep Speech 2 [62]

The architecture used in this work is shown in Figure 7. Our model consists of two 2D convolution layers with
a kernel size of (11, 41) and a stride size of (2, 2) for the first layer, a kernel size of (11, 21) and a stride size of
(1, 2) for the second layer, and 32 filters each. CNN layers are followed by 5 bidirectional GRU layers with 512
units for each layer and dropout layers of 0.5 rate. ReLU activation layers, and Batch normalization layers were
also used. Batch normalization helps stabilize the training of the model and reduces the chances of overfitting
by normalizing the activations of each layer to have zero mean and unit variance [63]. Rectified Linear Unit
(ReLU) layers are a type of activation function that introduce non-linearity to the model, to help the model to
learn more complex representations of the input data [64]. Then finally comes a dense layer containing 1024
neurons allows the model to learn interactions between different features, followed by a classification layer
consisting of 46 neurons, to classify the current input into the blank symbol used in the CTC algorithm or the
alphabet used in this work, which illustrated in Table 3.

Table 3 The alphabet used in this work

‫ح‬ ‫ج‬ ‫ث‬ ‫ت‬ ‫ب‬ ‫آ‬ ‫ا‬ ‫أ‬ ‫إ‬ ‫ء‬
‫ط‬ ‫ض‬ ‫ص‬ ‫ش‬ ‫س‬ ‫ز‬ ‫ر‬ ‫ذ‬ ‫د‬ ‫خ‬
‫ه‬ ‫ن‬ ‫م‬ ‫ل‬ ‫ك‬ ‫ق‬ ‫ف‬ ‫غ‬ ‫ع‬ ‫ظ‬
ِ ِ ِ ِ ‫ة‬ ‫ئ‬ ‫ى‬ ‫ي‬ ‫ؤ‬ ‫و‬
З Space ِ ِ ِ ِ

Adam optimizer with 1e-4 learning rate and CTC loss function were used to train the model and was
implemented using the TensorFlow library in Python and trained on the Google Colab platform7 which provides
access to an NVIDIA Tesla T4 GPU with 16 GB of memory.

7
Colaboratory: https://colab.research.google.com/
14
 Fig. 7 Our Proposed Solution Schema

6. Result & Discussion

As the Ar-DAD dataset includes recitations by a large number of readers at different speeds, we performed
many experiments to choose the best spectrogram extraction parameters that can be suitable for all speeds.
Choosing appropriate parameters, values of frame length, and hop size, in particular, is important to balance
the trade-off between time and frequency resolution when Fourier Transform is applied. In practice, the choice
of these parameters depends on the application and the characteristics of the signal being analyzed, and it's an
iterative process. The detailed experiments are shown in Table 4.

We used the most common and widely used metrics for measuring the performance of speech recognition
models, the Character Error Rate (CER), and the Word Error Rate (WER) metric. WER of 8.34% and CER of
2.42% were the best results we achieved.

The formula for Word Error Rate (WER) is:

𝑊𝐸𝑅 = (𝑆 + 𝐷 + 𝐼) / 𝑁 (9)

The formula for Character Error Rate (CER) is:

𝐶𝐸𝑅 = (𝑆 + 𝐷 + 𝐼) / 𝑀 (10)

Where: S, D and I are the number of substitutions, deletions, and insertions required to change the recognized
transcript into the reference transcript. N is the total number of words, and M is the total number of characters
in the reference transcript.

15
Table 4 Experiments conducted to select feature extraction parameters

Experiment FFT Hop Size Test set

Number WER CER
1 512 256 9.02% 2.64%
2 384 8.7% 2.45%

3 800 400 8.34% 2.77%

4 600 8.51% 2.42%

5 1024 512 10.7% 3.3%

6 768 11.5% 3.2%

As for comparing our results with the results of previous work, unfortunately, all the datasets used before are
private and this made it difficult for us to compare with others. Therefore, using a public dataset in this research
will solve the problem in the future so that researchers in this field can compare their work with us. However,
as we have shown before, the vast majority of the previous works consider the problem as a classification task
to detect the mispronunciation of Tajweed rules and some verses by samples containing wrong and correct
pronunciation. As for the few works that proposed solutions based on traditional ASR, as mentioned in the
literature review, all of them used data containing a few chapters of the Quran at the same speed and a few
verses by a small number of readers who recited those verses. So, their results cannot be compared with our
results as the dataset we used includes about 37 chapters of the Holy Quran in addition to a large number of
reciters, which reaches 30 reciters, who recited these chapters at different speeds and with different
applications of Tajweed rules. Table 5 shows a comparison between the real text and the text predicted by the
proposed model for some verses from the Ar-DAD dataset.

Table 5 Some samples that show the performance of the model

Actual Predicted
َّ ‫إ َّن الَّذين آمنوا وعملوا ال‬
‫صالحات أولئك هم خير البريَّة‬ َّ ‫إ َّن الَّذين آمنوا وعملوا ال‬
‫صالحات أولئك هم خير البريَّة‬
‫إ َّن الَّذين فتنوا المؤمنين والمؤمنات ث َّم لم يتوبوا فلهم عذاب جهنَّم‬ ‫إ َّن الَّذين فتنوا المؤمنين والمؤمنات ث َّم لم يتوبوا فلهم عذاب جهنَّم‬
‫ولهم عذاب الحريق‬ ‫ولهم عذاب الحريد‬
‫فأ َّما اْلنسان إذا ما ابتَله ربُّه فأكرمه ونعَّمه فيقول ربي أكرمن‬ ‫فأ َّما اْلنسان إذا ما ابتَله ربُّه فأكرمه ونعَّمه فيقول ربي أكرم‬
‫وجيء يومئذ بجهنَّم يومئذ يتذ َّكر اْلنسان وأنَّى له الذكرى‬ ‫وجيء يومئذ بجهنَّم يومئذ يتذ َّكر اْلنسان وأنَّى له الذكرى‬
‫لم يكن الَّذين كفروا من أهل الكتاب والمشركين منفكين حتَّى‬ ‫لم يكن الَّذين كفروا من أهل الكتاب والمشركين منفكين حتَّى‬
‫تأتيهم البينة‬ ‫تأتيهم البينة‬
‫صَلة‬ َّ ‫وما أمروا إ ََّّل ليعبدوا‬
َّ ‫ّللا مخلصين له الدين حنفاء ويقيموا ال‬ ‫وما أمروا إ ََّّل ليعبدوا الَّه مخلصين له الدين حنفاءن ويقيموا‬
َّ ‫ويؤتوا‬
‫الزكاة وذلك دين القيمة‬ َّ ‫صَلة ويؤتوا‬
‫الزكاة وذلك دين القيمة‬ َّ ‫ال‬
‫إ َّن الَّذين كفروا من أهل الكتاب والمشركين في نار جهنَّم خالدين‬ ‫إ َّن الَّذين كفروا من أهل الكتاب والمشركين في نار جهنَّم خالدين‬
‫فيها أولئك هم ش ُّر البريَّة‬ ‫فيها أولئك هم ش ُّر البريَّة‬

7. Conclusion
In conclusion, this research presents a novel end-to-end deep learning model for recognizing Holy Quran
recitation. Moreover, our proposed model provides the ability to give feedback to users about error type and
16
location so that users can have a better experience in correcting their errors in their learning journey. The
proposed solution consists of two main components: A CNN-Bidirectional GRU encoder uses the CTC loss
function and a character-based decoder. Using this end-to-end model allows us to get rid of alignment tools, so
reduce needed efforts, and improve performance. Our proposed model has been evaluated on a recently
published public data set (Ar-DAD), which contains about 37 chapters recited by 30 reciters. Ar-DAD dataset
used to train a model for recognizing Quran recitation has limitations. It only includes men reciters, making it
less reliable for recognizing recitations by women and children. Additionally, it only contains samples from one
recitation form, while there are ten approved forms, which may cause the model to incorrectly recognize
recitations with different forms. The model's performance was evaluated using word error rate (WER) and
character error rate (CER) as metrics, with the best results obtained being 8.34% WER and 2.42% CER. These
results demonstrate the effectiveness of the proposed model in recognizing the recitation of the Holy Quran
that outperforms previous related works. We hope that this paper will provide a baseline for fair comparisons
in this task, as it is based on a publicly available dataset that can be used by all researchers.

8. Declarations
Authors’ contributions
AAH performed the literature review, conducted the experiments, and wrote the manuscript.
KAJ took on a supervisory role and made a contribution to the conception and analysis of the work.
All authors read and approved the final manuscript.

Funding
The authors declare that they have no funding.

Data Availability
The data set used in this work is available at:

https://data.mendeley.com/datasets/3kndp5vs6b/3

Ethics approval and consent to participate

The authors Ethics approval and consent to participate.

Consent for publication

The authors consent for publication.

Conflicts of Interest
The authors declare that they have no competing interests.

References
[1] A. Khatib and N. Khan, “The Origins of the Variant Readings of the Qur’an | Yaqeen Institute for Islamic
Research,” Sep. 20, 2020. https://yaqeeninstitute.org/read/paper/the-origins-of-the-variant-
readings-of-the-quran#ftnt2 (accessed May 01, 2023).

[2] S. Nasser, The Transmission of the Variant Readings of the Qur’an. Leiden: Brill Publishers. Accessed:
May 01, 2023. [Online]. Available:

17
 https://books.google.nl/books?id=Kx7i2Y56WuYC&dq=aasim+qira%27ah&pg=PA57&redir_esc=y#v
=onepage&q&f=false

[3] M. Cavendish, Islamic Beliefs, Practices, and Cultures. 2010. Accessed: May 01, 2023. [Online].
Available:
https://books.google.nl/books?id=H_m14NlQQMYC&dq=Qira%CA%BCat+quran&pg=PA271&redir_e
sc=y#v=onepage&q&f=false

[4] M. Alghamdi, Arabic Phonetics and Phonology (in Arabic). 2015.

[5] H. Morsy, M. Shahin, N. Aljohani, M. Shoman, and S. Abdou, “Automatic speech attribute detection of
arabic language,” International Journal of Applied Engineering Research, vol. 13, no. 8, pp. 5633--5639,
2018.

[6] ‫ب‬. ‫مكريني‬, ‫معجم المصطلح الصوتي عند علماء التجويد‬. Dar Al Kotob Al Ilmiyah ‫دار الكتب العلمية‬, 2013. Accessed: Feb.
08, 2023. [Online]. Available:
https://books.google.com/books?id=WfZHDwAAQBAJ&pg=PA0#v=onepage&q&f=false

[7] H. A. Hassan, N. H. Nasrudin, M. N. M. Khalid, A. Zabidi, and A. I. Yassin, “Pattern classification in

recognizing Qalqalah Kubra pronuncation using multilayer perceptrons,” ISCAIE 2012 - 2012 IEEE
Symposium on Computer Applications and Industrial Electronics, pp. 209–212, 2012, doi:
10.1109/ISCAIE.2012.6482098.

[8] M. Al-Ayyoub, N. A. Damer, and I. Hmeidi, “Using Deep Learning for Automatically Determining
Correct Application of Basic Quranic Recitation Rules,” The International Arab Journal of Information
Technology, vol. 15, 2018, [Online]. Available:
https://www.researchgate.net/publication/326010238_Using_Deep_Learning_for_Automatically_Det
ermining_Correct_Application_of_Basic_Quranic_Recitation_Rules

[9] A. M. Alagrami and M. M. Eljazzar, “SMARTAJWEED Automatic Recognition of Arabic Quranic

Recitation Rules,” Computer Science & Information Technology (CS & IT), pp. 145–152, Dec.
2020, doi: 10.5121/CSIT.2020.101812.

[10] F. Ahmad, S. Z. Yahya, Z. Saad, and A. R. Ahmad, “Tajweed Classification Using Artificial Neural
Network,” 2018 International Conference on Smart Applications, Communications and Networking,
SmartNets 2018, Jul. 2018, doi: 10.1109/SMARTNETS.2018.8707394.

[11] R. M. Al-Khatib, M. Alshanaq, M. Barhoush, and K. Nahar, “AN EFFICIENT HOLY QURAN RECITATION
RECOGNIZER BASED ON SVM LEARNING MODEL,” Jordanian Journal of Computers and Information
Technology, vol. 6, pp. 392–414, Oct. 2020, doi: 10.5455/jjcit.71-1593380662.

[12] M. ibn I. Bukhā rī and M. Muhsin. Khan, “Ṣ aḥ īḥ al-Bukhā rī : the translation of the meanings of Sahih al-
Bukhari : Arabic-English,” Kazi Publications, vol. 6, 1997.

[13] H. Tabbal, W. El Falou, and B. Monla, “Analysis and implementation of a ‘Quranic’ verses delimitation
system in audio files using speech recognition techniques,” in 2006 2nd International Conference on
Information & Communication Technologies, IEEE, 2006, pp. 2979–2984. doi:
10.1109/ICTTA.2006.1684889.

[14] B. Putra, B. Atmaja, and D. Prananto, “Developing Speech Recognition System for Quranic Verse
Recitation Learning Software,” IJID (International Journal on Informatics for Development), vol. 1, no. 2,
p. 14, Dec. 2012, doi: 10.14421/ijid.2012.01203.

18
[15] M. Ab. Rahman, I. A. A. Kassim, T. Ab. Rahman, and S. Z. M. Muji, “Development of Automated Tajweed
Checking System for Children in Learning Quran,” Evolution in Electrical and Electronic Engineering,
vol. 2, no. 1, pp. 165–176, May 2021, doi: 10.30880/eeee.2021.02.01.019.

[16] A. Muhammad, Z. u Qayyum, W. M. M., S. Tanveer, M.-E. A.M., and A. Z. Syed, “E-Hafiz: Intelligent
System to Help Muslims in Recitation and Memorization of Quran,” Life Sci J, vol. 9, pp. 534–541, 2012,
[Online]. Available: https://www.researchgate.net/publication/291155414_E-
Hafiz_Intelligent_System_to_Help_Muslims_in_Recitation_and_Memorization_of_Quran

[17] R. A. Rajagede and R. P. Hastuti, “Al-Quran recitation verification for memorization test using Siamese
LSTM network,” Communications in Science and Technology, vol. 6, no. 1, pp. 35–40, Jul. 2021, doi:
10.21924/CST.6.1.2021.344.

[18] H. M. A. Tabbaa and B. Soudan, “Computer-Aided Training for Quranic Recitation,” Procedia Soc Behav
Sci, vol. 192, pp. 778–787, Jun. 2015, doi: 10.1016/J.SBSPRO.2015.06.092.

[19] A. A. Al-Bakeri, “ASR for Tajweed Rules: Integrated with SelfLearning Environments,” International
Journal of Information Engineering and Electronic Business, vol. 9, no. 6, pp. 1–9, Nov. 2017, doi:
10.5815/IJIEEB.2017.06.01.

[20] “Language Modeling Toolkit.” http://www.cs.cmu.edu/~dorcas/toolkit_documentation.html

(accessed Jan. 18, 2023).

[21] I. K. Tantawi, M. A. M. Abushariah, and B. H. Hammo, “A deep learning approach for automatic speech
recognition of The Holy Qur’ān recitations,” Int J Speech Technol, vol. 24, no. 4, pp. 1017–1032, Dec.
2021, doi: 10.1007/s10772-021-09853-9.

[22] A. A. M. Alqudah, M. A. M. Alshraideh, and A. A. S. Sharieh, “Arabic Disordered Speech Phonetic

Dictionary Generator for Automatic Speech Recognition,” J Theor Appl Inf Technol, vol. 98, no. 4, pp.
571–586, 2020.

[23] A. Stolcke, “Srilm — An Extensible Language Modeling Toolkit,” Proceedings of the 7th International
Conference on Spoken Language Processing (ICSLP 2002), vol. 2, Jan. 2004.

[24] J. Hui, “Speech Recognition — ASR Model Training,” Medium, Oct. 01, 2019. https://jonathan-
hui.medium.com/speech-recognition-asr-model-training-90ed50d93615 (accessed Jan. 29, 2023).

[25] Y. Seddiq, A. Meftah, M. Alghamdi, and Y. Alotaibi, “Reintroducing KAPD as a Dataset for Machine
Learning and Data Mining Applications,” Proceedings - UKSim-AMSS 2016: 10th European Modelling
Symposium on Computer Modelling and Simulation, pp. 70–74, May 2017, doi: 10.1109/EMS.2016.022.

[26] “Arabic Speech Corpus | Kaggle.”

https://www.kaggle.com/datasets/bc297d8ca0753cd21cdcacd7bd324c0c607361a14471c801f09b0
28a1ecb098e (accessed Jan. 19, 2023).

[27] S. Wang and G. Li, “Overview of end-to-end speech recognition,” J Phys Conf Ser, vol. 1187, no. 5, p.
052068, Apr. 2019, doi: 10.1088/1742-6596/1187/5/052068.

[28] W. M. Muhammad, R. Muhammad, A. Muhammad, and A. M. Martinez-Enriquez, “Voice Content

Matching System for Quran Readers,” in 2010 Ninth Mexican International Conference on Artificial
Intelligence, IEEE, Nov. 2010, pp. 148–153. doi: 10.1109/MICAI.2010.11.

[29] A. Hussein, S. Watanabe, and A. Ali, “Arabic speech recognition by end-to-end, modular systems and
human,” Comput Speech Lang, vol. 71, p. 101272, Jan. 2022, doi: 10.1016/j.csl.2021.101272.

19
[30] A. Ali et al., “The MGB-2 challenge: Arabic multi-dialect broadcast media recognition,” 2016 IEEE
Workshop on Spoken Language Technology, SLT 2016 - Proceedings, pp. 279–284, Feb. 2017, doi:
10.1109/SLT.2016.7846277.

[31] A. Ali, S. Vogel, and S. Renals, “Speech recognition challenge in the wild: Arabic MGB-3,” 2017 IEEE
Automatic Speech Recognition and Understanding Workshop, ASRU 2017 - Proceedings, vol. 2018-
January, pp. 316–322, Jan. 2018, doi: 10.1109/ASRU.2017.8268952.

[32] A. Ali et al., “The MGB-5 Challenge: Recognition and Dialect Identification of Dialectal Arabic Speech,”
2019 IEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2019 - Proceedings, pp.
1026–1033, Dec. 2019, doi: 10.1109/ASRU46091.2019.9003960.

[33] A. Ahmed, Y. Hifny, K. Shaalan, and S. Toral, “End-to-End Lexicon Free Arabic Speech Recognition
Using Recurrent Neural Networks,” in Computational Linguistics, Speech and Image Processing for
Arabic Language, WORLD SCIENTIFIC, 2018, pp. 231–248. doi: 10.1142/9789813229396_0011.

[34] Y. Belinkov, A. Ali, and J. Glass, “Analyzing Phonetic and Graphemic Representations in End-to-End
Automatic Speech Recognition,” Proceedings of the Annual Conference of the International Speech
Communication Association, INTERSPEECH, vol. 2019-September, pp. 81–85, Jul. 2019, doi:
10.48550/arxiv.1907.04224.

[35] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public
domain audio books,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), IEEE, Apr. 2015, pp. 5206–5210. doi: 10.1109/ICASSP.2015.7178964.

[36] A. Rousseau, P. Deléglise, and Y. Estève, “TED-LIUM: an Automatic Speech Recognition dedicated
corpus,” in Proceedings of the Eighth International Conference on Language Resources and Evaluation
(LREC’12, Istanbul, Turkey: European Language Resources Association (ELRA, 2012, pp. 125–129.
[Online]. Available: http://www.lrec-conf.org/proceedings/lrec2012/pdf/698_Paper.pdf

[37] H. A. Alsayadi, A. A. Abdelhamid, I. Hegazy, and Z. T. Fayed, “Arabic speech recognition using end‐to‐
end deep learning,” IET Signal Processing, vol. 15, no. 8, pp. 521–534, Oct. 2021, doi:
10.1049/sil2.12057.

[38] S. Watanabe et al., “ESPnet: End-to-End Speech Processing Toolkit,” Proceedings of the Annual
Conference of the International Speech Communication Association, INTERSPEECH, vol. 2018-
September, pp. 2207–2211, Mar. 2018, Accessed: Oct. 13, 2022. [Online]. Available:
http://arxiv.org/abs/1804.00015

[39] Y. Wang et al., “Espresso: A Fast End-to-End Neural Speech Recognition Toolkit,” in 2019 IEEE
Automatic Speech Recognition and Understanding Workshop (ASRU), IEEE, Dec. 2019, pp. 136–143. doi:
10.1109/ASRU46091.2019.9003968.

[40] M. Lataifeh and A. Elnagar, “Ar-DAD: Arabic diversified audio dataset,” Data Brief, vol. 33, p. 106503,
Dec. 2020, doi: 10.1016/j.dib.2020.106503.

[41] O. Abdel-Hamid, A. Mohamed, H. Jiang, and G. Penn, “Applying Convolutional Neural Networks
concepts to hybrid NN-HMM model for speech recognition,” in 2012 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), IEEE, Mar. 2012, pp. 4277–4280. doi:
10.1109/ICASSP.2012.6288864.

[42] T. N. Sainath, A. Mohamed, B. Kingsbury, and B. Ramabhadran, “Deep convolutional neural networks
for LVCSR,” in 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE,
May 2013, pp. 8614–8618. doi: 10.1109/ICASSP.2013.6639347.

20
[43] H. Soltau, G. Saon, and T. N. Sainath, “Joint training of convolutional and non-convolutional neural
networks,” in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
IEEE, May 2014, pp. 5572–5576. doi: 10.1109/ICASSP.2014.6854669.

[44] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural networks,” IEEE Transactions on Signal
Processing, vol. 45, no. 11, pp. 2673–2681, 1997, doi: 10.1109/78.650093.

[45] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural
Networks,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2012.
Accessed: Feb. 08, 2023. [Online]. Available: http://code.google.com/p/cuda-convnet/

[46] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image
Recognition,” 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track
Proceedings, Sep. 2014, doi: 10.48550/arxiv.1409.1556.

[47] Y. Zhang et al., “Towards End-to-End Speech Recognition with Deep Convolutional Neural Networks,”
Proceedings of the Annual Conference of the International Speech Communication Association,
INTERSPEECH, vol. 08-12-September-2016, pp. 410–414, Jan. 2017, doi: 10.48550/arxiv.1701.02720.

[48] C. Olah, “Neural Networks, Types, and Functional Programming,” GitHub Pages, Sep. 03, 2015.
https://colah.github.io/posts/2015-09-NN-Types-FP/ (accessed Oct. 12, 2022).

[49] S. Hochreiter, “The Vanishing Gradient Problem During Learning Recurrent Neural Nets and Problem
Solutions,” International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 06, no.
02, pp. 107–116, Apr. 1998, doi: 10.1142/S0218488598000094.

[50] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies with gradient descent is
difficult,” IEEE Trans Neural Netw, vol. 5, no. 2, pp. 157–166, Mar. 1994, doi: 10.1109/72.279181.

[51] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Comput, vol. 9, no. 8, pp. 1735–
1780, Nov. 1997, doi: 10.1162/neco.1997.9.8.1735.

[52] K. Cho, B. van Merriënboer, D. Bahdanau, and Y. Bengio, “On the Properties of Neural Machine
Translation: Encoder-Decoder Approaches,” Proceedings of SSST 2014 - 8th Workshop on Syntax,
Semantics and Structure in Statistical Translation, pp. 103–111, Sep. 2014, doi:
10.48550/arxiv.1409.1259.

[53] M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio, “Light Gated Recurrent Units for Speech
Recognition,” IEEE Trans Emerg Top Comput Intell, vol. 2, no. 2, pp. 92–102, Apr. 2018, doi:
10.1109/TETCI.2017.2762739.

[54] Y. Su and C.-C. J. Kuo, “On extended long short-term memory and dependent bidirectional recurrent
neural network,” Neurocomputing, vol. 356, pp. 151–161, Sep. 2019, doi:
10.1016/j.neucom.2019.04.044.

[55] A. Perambai, “A deep dive into the world of gated Recurrent Neural Networks: LSTM and GRU,”
Medium, Aug. 02, 2019. https://medium.com/analytics-vidhya/lstm-and-gru-a-step-further-into-the-
world-of-gated-rnns-99d07dac6b91 (accessed Oct. 13, 2022).

[56] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification,” in

Proceedings of the 23rd international conference on Machine learning - ICML ’06, New York, New York,
USA: ACM Press, 2006, pp. 369–376. doi: 10.1145/1143844.1143891.

21
[57] H. Scheidl, S. Fiel, and R. Sablatnig, “Word beam search: A connectionist temporal classification
decoding algorithm,” Proceedings of International Conference on Frontiers in Handwriting Recognition,
ICFHR, vol. 2018-August, pp. 253–258, Dec. 2018, doi: 10.1109/ICFHR-2018.2018.00052.

[58] A. Maas, Z. Xie, D. Jurafsky, and A. Y. Ng, “Lexicon-free conversational speech recognition with neural
networks,” in Proceedings of the 2015 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, 2015, pp. 345–354.

[59] Y. Kim, Y. Jernite, D. Sontag, and A. M. Rush, “Character-aware neural language models,” in Thirtieth
AAAI conference on artificial intelligence, 2016.

[60] M. X. Chen et al., “The Best of Both Worlds: Combining Recent Advances in Neural Machine
Translation,” ACL 2018 - 56th Annual Meeting of the Association for Computational Linguistics,
Proceedings of the Conference (Long Papers), vol. 1, pp. 76–86, Apr. 2018, doi:
10.48550/arxiv.1804.09849.

[61] I. Sutskever, O. Vinyals, and Q. V Le, “Sequence to Sequence Learning with Neural Networks,” in
Advances in Neural Information Processing Systems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence,
and K. Q. Weinberger, Eds., Curran Associates, Inc., 2014. [Online]. Available:
https://proceedings.neurips.cc/paper/2014/file/a14ac55a4f27472c5d894ec1c3c743d2-Paper.pdf

[62] D. Amodei et al., “Deep Speech 2: End-to-End Speech Recognition in English and Mandarin,” 33rd
International Conference on Machine Learning, ICML 2016, vol. 1, pp. 312–321, Dec. 2015, doi:
10.48550/arxiv.1512.02595.

[63] S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing
Internal Covariate Shift,” 32nd International Conference on Machine Learning, ICML 2015, vol. 1, pp.
448–456, Feb. 2015, doi: 10.48550/arxiv.1502.03167.

[64] V. Nair and G. E. Hinton, “Rectified Linear Units Improve Restricted Boltzmann Machines,” 2010.

22