Masry: A Text-to-Speech System for the Egyptian Arabic

Ahmed Hammad Azab

, Ahmed B. Zaky

2,4 b

, Tetsuji Ogawa

3 c

and Walid Gomaa

1,5 d

Computer Science and Engineering, Egypt-Japan University of Science and Technology, Alexandria, Egypt

Computer Science and Information Technology Programs (CSIT),

Egypt Japan University of Science and Technology, Egypt

Department of Communications and Computer Engineering, Waseda University, Tokyo, Japan

Shoubra Faculty of Engineering, Benha University, Benha, Egypt

Faculty of Engineering, Alexandria University, Alexandria, Egypt

Keywords:

Natural Language Processing, Text-To-Speech, Egyptian Arabic.

Abstract:

This paper presents the improvement and evaluation of Masry, an end-to-end system planned to synthesize

Egyptian Arabic speech. The proposed approach leverages the capable Tacotron speech synthesis models,

counting Tacotron1 and Tacotron2, and integrated with progressed vocoders – Grifﬁn-Lim f or Tacotron1 and

HiFi-GAN for Tacotron2. By synthesizing waveforms from mel-spectrograms, Masry offers a comprehensive

solution for generating natural and expressive Egyptian Arabic speech. To train and validate our system, we

construct a dataset including a male speaker describing standard composing pieces and news content in Egyp-

tian Arabic. The sampling rate of recorded data is 44100 Hz, guaranteeing constancy and richness within the

synthesized speech output. The execution of our framework was fastidiously assessed t hrough different mea-

surements, with a speciﬁc center on the Mean Opinion Score (MOS). The experimental results demonstrated

the prevalence of Tacotron2 over Tacotron1, yielding a MOS of 4.48 compared to 3.64. This emphasizes

the system’s capacity to capture and duplicate the nuances of Egyptian Arabic speech more effectively. Be-

sides, The assessment extended to include fundamental measurements such as word and character error r at es

(WER and CER). These metrics give a quantitative appraisal of the precision and exactness of the synthesized

speech.

1 INTRODUCTION

Text-to-Speech (TTS) technology has become a crit-

ical ﬁeld of research and development. It aims at

transforming written text into spoken words, enabling

applications like voice assistants, audiobooks, acces-

sibility tools, and language learning platforms with

the potential to revolutionize human-computer inter-

actions.(Youn g et al., 2018)

Arabic stands as one of the most extensively spo-

ken languages, serving as the mother tongue for over

200 million individuals (Versteegh, 2014), and the

largest Semitic languag e. It is composed of two

main dialects: Standard Arabic and Dialec ta l Ara-

bic. While Modern Standar d Arabic (MSA) is the for-

mal linguistic standard, Dialectal Arabic represents

https://orcid.org/0009-0007-2461-1040

https://orcid.org/0000-0002-3107-5043

https://orcid.org/0000-0002-7316-2073

https://orcid.org/0000-0002-8518-8908

the daily spoken variation, exhibiting signiﬁcant dif-

ferences within and across countries(Habash, 202 2).

However, the literature ha s bee n dominate d by

TTS systems for English, resulting in a gap in de-

veloping TTS systems for less commonly spo ken

low-resource languages and dialects, including Egyp -

tian Arabic (Fahmy et al., 2020). As a widely spo-

ken dialect with unique regional variations and in-

formal characteristics (Abdel-Massih, 2011), Egyp-

tian Arab ic requires dedicate d attention to its lin-

guistic nuances and cultural relevance. Hence, de-

veloping an efﬁcient and accurate TTS system for

Egyptian Arabic is paramoun t to en hancing acc essi-

bility and communication for its users. The current

work aims to address this imperative by presenting a

novel approach for building a high-quality TT S sys-

tem speciﬁcally tailored to the unique char acteristics

of Egyptian Arabic, drawing insights from existing re-

lated work in both Arabic and English TTS systems

(Habash, 2022) .

Azab, A., Zaky, A., Ogawa, T. and Gomaa, W.

Masry: A Text-to-Speech System for the Egyptian Arabic.

DOI: 10.5220/0012244300003543

In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 2, pages 219-226

ISBN: 978-989-758-670-5; ISSN: 2184-2809

219

Moreover, diacritic and gemination signs, repre-

senting short vowels and consonan t doubling, respec-

tively, play a crucial role in corr ectly pronouncing

Arabic and its various dialects. However, these signs

are often omitted in written texts as most Arab read-

ers are accustomed to inferring them from the con-

text (Habash, 2022). So, this absence poses signiﬁ-

cant difﬁculties fo r TTS systems aiming to accurately

represent the diverse pronunciations in Arabic.

The lack of Egyptian datasets poses a signiﬁcant

challenge for training Text-to-Speech (TTS) models.

As TTS technology strives for natural and accurate

speech synthesis, it heavily relies on large and di-

verse datasets in various langu ages, includin g Egyp-

tian Arabic. Unfortu nately, the scarcity of high-

quality, annotated data in this speciﬁc d ia le c t hinders

the development of accurate TTS systems that can ef-

fectively mimic the unique ch aracteristics of Egyptian

Arabic speech (Baali et al., 2023). With out sufﬁcient

training data, TTS mod e ls m ay struggle to capture

the nuances of pronunciation, intonation, and linguis-

tic variations speciﬁc to the Egyptian dialect, leading

to less authentic and less intelligible speech synthe-

sis. Addressing this issue requires collaborative ef-

forts to collect and curate more Egyptian datasets, fos-

tering the advancement of TTS technology to cater to

a broader linguistic landscape. Our contributions are

as follows:

1. We present a dataset for a male (named, Ashraf)

narrating general writing and news in Egyptian

Arabic.

2. We present “Masry” an end-to-end text-to-speech

system for Egyptian A rabic.

The paper is structured as fo llows: In Section 1,

we p rovide an introduction to the study’s objectives

and scope. Section 2 is dedicated to the review of

related literature and prior works in the ﬁeld. Mov-

ing forward, Sectio n 3 elaborates on the system ar-

chitecture, delv ing into each phase within this fra me-

work. Expanding upon the experimental setup, Sec-

tion 4 elucidates the conducted experiments involving

the models. In Section 5, we present the outcomes

of th ese experiments and engage in a comprehensive

discussion ther eof. Finally, Section 6 encapsu late s the

paper with a conclusion summarizing our ﬁndings and

outlining potential avenues for future research.

2 RELATED WORK

Arabic text-to-speech synthesis is one o f AI’s NLP

challenges. Ma ny attempts have bee n tried to make

systems that ca n overcome the problem of artiﬁcial

voice to create a m ore human natural voice. Previous

works have been done in this a rea, but most models

are made for the English language. Some models have

been applied to Arabic withou t focusing on speciﬁc

dialects. (Habash, 2022).

2.1 English TTS Models

In English, many TTS models have bee n developed.

The most famous one is Tacotron. Tacotron is an

end-to-end generative text-to- speech model that can

directly synthesize speech from the text with a sim-

ple waveform synthesis module. Its highest achieved

MOS score (Mean Opinion Score) is 3.82 (Wang

et al., 2017).

The authors in (Ren et al., 2019) proposed a

novel method (FastSpeech) using a feed-forward net-

work based on Transformer that generates a mel-

spectrogram in parallel for TTS. They extract at-

tention alignments from an encoder-decoder-based

teacher model for phoneme duration pre diction. A

length regulator uses it to exp and the source phoneme

sequence to match the length of the target mel-

spectrogram sequence for producing parallel mel-

spectrograms. The co nducted experiments on the

LJSpeech dataset (I to and Johnson, 2017) reveal that

their parallel model achieves a compar able level of

speech quality to autoregressive models.

2.2 Arabic TTS Models

Some TTS models have been developed for the

Arabic lan guage. Ossam et a l. (Abdel-Hamid

et al., 20 06) employed an HMM-based (hidde n

Markov model) ap proach to en hance synthesized Ara-

bic speech. Their methodology used a statistical

model to generate Arabic speech parameters like

spectrum, fundamental frequen cy (F0), and d uration

of phonemes. They also inco rporated a m ulti-band

excitation m odel and utilized samples from the spec-

tral envelope as spectral parameters.

Zangar et al. ( Imene et al., 2018) focused on utiliz-

ing Deep Neural Networks (DNN) for duration mod-

eling in Arabic speech synthesis. They compared

HMM-based and deep neural network D N N-based

duration modeling of various architectures to mini-

mize roo t mean square prediction error (RMSE). The

study concluded that using their DNN for modeling

duration outperformed both the HMM -based model-

ing from the HTS toolkit and the DNN-based mo del-

ing from the MERLIN toolkit.

In (Fahmy et al., 2020), the authors proposed a

transfer learning end-to-e nd Modern Standard Arabic

(MSA) TTS deep architecture. Their work presents

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

220

how they g e nerate high -quality, natu ral, and human-

like Arabic spee c h th at uses an end-to-end neural deep

network architectur e . The approach is built upon a

limited corpus of text and audio pairs, encompass-

ing a relatively small compilation of recorded audio,

amounting to 2.41 hours. Notably, it demonstrates

the successful utilization of English character embed-

dings, even when employing diacritic Arabic charac-

ters as input. The study further expounds on the pre-

processing techniques applied to these audio samples,

elucidating the strategies used to optim iz e outcomes.

The authors in (Abdelali et al., 2022) proposed

an end-to-end TTS system for Arabic. They called

it the NatiQ system. Their speech synthesizer uses

an encoder-decode r architecture with attentio n. They

used the Tacotron model ( Tacotr on1 a nd Tacotron2)

and transformer mod e l to generate mel-spe c trograms

from ch a racters. They used WaveRNN vocoder with

Tacotron1, WaveGlow vocoder with Tacotron2, an d

ESPnet transformer with the parallel wavegan to

synthesize waveforms from the spectrograms. Two

voices, male and female, are used. The authors

achieved 4.21 ( MOS) for the female voice and 4.40

for the male voice.

3 SYSTEM ARCHITECTURE

The system, Masry, is structured into three key el-

ements, depicted in Fig.1. The initial stage in-

volves data preprocessing to reﬁne input data. Sub-

sequently, a text-to- mel-spectrogram model ge nerates

mel-spectrum output. Finally, the mel-spectrum is

processed by a Vocoder to generate corresponding au-

dio. Additionally, we elaborate on the dataset col-

lected, used for training purposes.

3.1 Dataset

In this section, we will discuss the dataset we col-

lected and its characteristics.

EGYARA-23 Dataset

We collected a dataset and called it EGYARA-23. The

EGYARA-23 dataset comprises recordin gs of a male

speaker named Ashraf, who narrates general co nver-

sations and news in Egyptian Arabic. The dataset

spans 20 .5 hour s and con sists of 105,329 words and

32,716 segments. On average, each segment has a

duration of 8 seconds, and the recordings maintain a

quality of 44.1 KHz.

We intended to create a comprehensive da ta set en-

compassing a wide rang e of Eg yptian Arabic words,

considerin g that this dialect contains numerous words

not present in Modern Standard Arabic (MSA). To en-

sure authenticity, we transcribed the dataset in Egyp-

tian Arabic as used in everyda y language, reﬂecting

the actual dialect. Consequently, each segment is ac -

compan ie d by its respective transcript. Table 1 dis-

plays a selection of MSA words and their equivalents

in Egyptian Arabic.

Table 1: Difference between words in MSA and Egyptian

Arabic.

Egyptian Arabic MSA English







J@







K@ Wait

hðQK



X@ Go



àA



«

Ég



Because



PðA«



P@ Want



ÊªÓ







CË

Sorry



HAë ú





æ¢«@

Give me

3.2 Preprocessing

In this sectio n, we will delve into the preprocessing

steps, which include diacritization, segmentation, and

phone tiza tion.

3.2.1 Diacritization

Egyptian Arabic diacritization encompasses two

types of vowels: long vowels, explicitly indicate d in

the text, and short vowels (diacritics), which are of-

ten omitted in modern writing (Ha bash, 2022), re-

lying on readers’ co ntextual understand ing (Abdelali

et al., 2022). Accurately restoring these diacritics is

paramount for human comprehension an d machine-

based pronunciation of Arabic words. To address this,

we employed Camel Tools (Obeid et al., 2020) as a

diacritization tool for Egyptian Arabic, attaining an

accuracy exceedin g 90%. However, to ensure meticu-

lousness, an expe rt in Arabic linguistics reviewed the

automatically diacritized data, particularly in intricate

cases such as named entities and foreign words, which

can present cha llenges even to native speakers. Dia-

critization holds signiﬁcant importance in accurately

pronouncing words, thus rendering it a crucial step

in the preprocessing phase. In Table 2, you can see

some Egyptian Arabic words before and after the di-

acritiazation of Egyptian Arabic.

3.2.2 Segmentation

Given the constraints of neural architectures in pro-

cessing lengthy audio samples (Shen et al., 2018),

We developed a semi- automated procedure for data

collection, where the dataset is partitioned into ap-

Masry: A Text-to-Speech System for the Egyptian Arabic

221

Vocoder

Preprocessing

(Diacritization/

Segmentation

/Phonetising)

Text to Mel-

Spectrogram Model

Input Text

Output Audio

Mel-

Spectrum

Processed

Text

(Phonems)

Speech Synthesize System

Speech Synthesize Model

Figure 1: Masry Architecture.

Table 2: Before and after diacritization of samples in E gyp-

tian Arabic.

Before After English







J@















@



Wait

hðQK



hð







àA



«





àA









« Because



PðA«





Pð



Want



ÊªÓ





Ê







Sorry



HAë







Give me

proxim ately 8-second frames. This segmentation pro-

cess involves meticulous attention to maintaining sen-

tence coherence while preservin g the overall contex-

tual ﬂow and prosody. Typically, extended pauses

between segments are utilized as reliable indicators

for segmentation, with exceptions being made when

extended pauses are followed by r elevant context or

supplementary conten t that remains part of the sen-

tence (Abdelali et al., 2022). This underscores the

signiﬁcance of a n intricate a pproach to segmentation

to uphold the precision of audio data representation

and its coherent continuity. Employing our semi-

automated pro cess, we address this challenge during

data collection, resulting in a dataset where each sen-

tence is inherently segmented with established start

and end times. Con sequently, the audio trimming pro-

cess is automated to align with each sentence’s corre-

sponding segments. We ensur e the entirety of each

sentence during d a ta collection, e nsuring that no per-

tinent context or supplementary content is left unre-

solved within the sentence.

3.2.3 Phonetization

Phonetization converts textual input into its cor-

respond ing phonetic representation, mapping each

graphe me to its phoneme. In TTS systems, p honetiza-

tion is crucial to accurately synthesize speech by en-

suring correct pronunciation, intonation, and r hythm.

For languages like Arabic, with intricate vowel pat-

terns and diacritics, phonetization is particularly im-

portant to cap ture pronunciation nuances. It e nables

TTS systems to generate natural an d con textua lly

appropriate speech output, creating a more expres-

sive and high-quality au ditory experience (El-Imam,

2004). In this phase, o ur focus shifts towards convert-

ing the transcripted text into a p honeme representa-

tion, which has already u ndergone diacritization and

segmentation. This transfor mation occu rs through a

two-step procedure as you can see in 2. Initially, we

transcribe the text into Buckwalter transliteration for-

mat. Subsequently, the output from the translitera-

tion step is subjected to further processing to obtain

phone me characters. To accomplish this task effec-

tively, we leverage the Arabic phonetiser tool devel-

oped by Nawar Halabi (Halabi, 2016). These sequen-

tial steps are essential in preparing the data for subse-

quent m odel input, optimizing our research approach

in Egyptian Arabic lang uage processing. you can see

before and after Phonetization in 3.

Buckwalter

Converter

Phonetizer PhonemesArabic Text

Figure 2: Phonetization Process.

Table 3: Difference between Arabic Text, Buckwalter

Transliteration and Phonemes Characters.

Phonemes Buckwalter Arabic Text

E a $ aa n Ea$aAno





àA









E aa w i1 z EaAwizo





Pð



m a E l i1 $ maEoli$o





Ê







h a a t haAto







< i0 s t a nn ii0 Aisotan∼iy ú















@



ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

222

3.3 Speech Synthesis Model

This sectio n presents an overview of the synthe-

sis model’s architecture. The synthesizer comprises

an encoder-decoder model and a vocoder which is

pivotal in gener ating the desired waveforms. The

encoder-decoder module is respo nsible f or converting

the preprocessed text into a mel-spec trum represen-

tation. Conversely, the vocoder transforms the mel-

spectrum representation into the corresponding wave-

form. To comprehensively explore the effectiveness

of the proposed approach, we experimented with two

distinct models and vocoders.

Tacotron1, (Wang et al., 2017), adopts an RNN

sequence-to-sequen ce architecture comprising three

main comp onents in Fig.3: an encoder, an attention-

based decoder, and a post-processing module. The

encoder ta kes text inp ut in the form o f char a cters

and transforms it into a mel-spectrogram represen-

tation. Subsequen tly, the post-processing module

utilizes this mel-spectrogram to generate the corre-

sponding waveform. The encoder in Tacotr on1 uti-

lizes a CBHG-based approach, wh ic h involves a bank

of 1-D convolutio nal ﬁlters, fo llowed by highway net-

works and a bidirectional gated recurrent unit (GRU).

Tacotron1 employs the Grifﬁn-Lim algorithm on

top of the generated me l-spectrograms to complete

the speech synthesis pro cess. The Grifﬁn-Lim algo-

rithm is a famou s vocoder used in spee c h synthesis,

including Tacotron1. It converts a mel-spectrogram

back into a time-domain waveform. It works itera-

tively, reﬁning a rand om waveform estimate to match

the target mel-spectr ogram. It retains phase informa-

tion for better waveform reconstruction by altern a ting

between time and frequency domains. This computa-

tionally efﬁcient approach strikes a balance betwe e n

quality and efﬁciency, making it widely used in real-

time speech synthesis.

Decoder

RNN

Decoder

RNN

Pre-netPre-net Pre-net

Attention

RNN

Attention

RNN

Attention

RNN

Decoder

RNN

Griffin-Lim reconstruction

Attention

Pre-net

CBHG

Character embeddings

Attention is applied

to all decoder steps

Linear-scale

spectrogram

Seq2seq target

with r=3

<GO> frame

Figure 3: Tacotron 1 architecture (Wang et al ., 2017).

Tacotron2 (Sh en et al., 2018) is an advanced text-

to-speech (TTS) model. In Fig. 4 consisting of a text

encoder and a spectro gram genera tor. The text en-

coder processes the input text, captur ing its linguis-

tic features and context, and produces a ﬁxed-size

representation. The spec trogram generator, equipped

with attention mechanisms and typically based on re-

current neural networks (RNNs), takes the ﬁxed-size

text representation as input and gene rates mel spectro-

grams, which represent the spectral conten t of audio

over time. During training, Tacotro n2 le arns to alig n

the inpu t text with corresponding mel spectrograms

using attention mechanisms, enabling it to generate

accurate and expressive mel spectrogra ms. On the

other hand, HiFi-GAN (Kong et al., 2020), the high-

ﬁdelity generative adversarial network vocoder, is de-

signed to convert mel spectro grams into high-quality

audio waveforms. By utilizing a GAN architecture

with a generator and a discriminator, HiFi-GAN is

trained to synthesize realistic and natural-sounding

speech from mel spec trograms. T he generator pro-

duces the audio waveforms while the discriminator

tries to distinguish between real and generated audio,

leading to adversarial tra ining that enhances th e qual-

ity of the generate d speech. With the combination of

Tacotron2 a nd HiFi-GAN, the TTS system can gener-

ate human-like speech by ﬁrst generating mel spec-

trogram s tha t retain essential char acteristics of the

audio a nd then utilizing the high-ﬁde lity vocoder to

transform these spectrograms into realistic and high-

quality speech waveforms.

Figure 4: Tacotron 2 architecture (Shen et al., 2018) with

modiﬁcation HiFi-GAN vocoder.

4 EXPERIMENTS

To evaluate the performance of each model, we per-

formed computational experiments with automatic

and manual (subjective) evaluations.

Masry: A Text-to-Speech System for the Egyptian Arabic

223

4.1 Training Phase

In the training proced ure, we adopt a two-step ap-

proach . Firstly, we tra in the feature prediction net-

work independently to predict certain features from

phone mes input. Subsequ ently, we proceed to train a

HiFi-GAN separately, utilizing the outputs ge nerated

by the feature prediction n etwork as its input. This

two-step pr ocess allows us to effectively leverage the

predictions made by the ﬁrst network and enhance the

overall performanc e of the HiFi-GAN. The training

process involved setting hyperpar ameters, such as a

batch size of 32, and conducting 1000 epochs. Ini-

tially, we preprocessed the input text by converting

it into phonemes. Subsequently, the trainin g proce-

dure commenced, during which we took checkpoints

at every 1000 steps. These checkpoints were used to

test and monitor the trainin g progress to ensur e effec-

tiveness an d efﬁciency. The computing environment

utilized an Intel Xeon 6 230R @ 2.1 G H z (52 CPUs)

with NVIDIA Quadro RTX500 0.

4.2 Testing Phase

To be able to test the models we build a testing dataset

called EGYARA-TEST that shares the same charac-

teristics as the EGYARA-23. However, EGYARA-

TEST encompasses a duration of 1 hour and com-

prises 1003 segments. This new dataset was explic-

itly designed to thorough ly test the performance of the

models on a mo re extensive and d iverse set of audio

samples. By using this dataset, we aimed to evaluate

how well the models perform and assess their robust-

ness and generalization capabilities across a broader

range of speech segments.

For mode l evaluation, we employ two types of as-

sessment: automatic evaluation a nd manual (su bjec-

tive) evaluation.

4.2.1 Automatic Evaluation

In this study, we employed an Egyptian Arabic Au-

tomatic Speech Recognition (ASR) system developed

by (Alyafeai, 2022) to decode the audio ﬁles gener-

ated by our Text-to- Speech (TTS) models. To com-

pare the generated transcripts with the input sentences

used for TTS output, we adapted the refere nce orig-

inal text to m atch th e unvowelized output from the

ASR system for a fair comparison . Standa rd evalua-

tion metrics, inc luding Word Error Rate (WER) and

Character Error Rate (CER), were u sed to assess the

TTS model’s perf ormance. Word Error Rate (WER)

is a widely used me tric for automatic speech recogni-

tion perfo rmance. It tackles challenges arising from

variable seq uence leng ths by employing Levenshtein

distance at the word level. WER facilitates system

compariso ns and improvements assessment, though

it lacks speciﬁcity on error typ e s. Addressing this,

dynamic strin g alignment aligns recog nized and ref-

erenced word sequences. The power law theory ex-

amines pe rplexity’s correlation w ith WER, shedding

light on language model complexity’s impact on error

rates (Morris et al., 2004). Word Erro r Rate (WER)

can subsequen tly be calculated as:

W ER = (S + D + I)/N = (S + D + I)/(S + D + C)

(1)

where S is the number of substitutions, D is the num-

ber of deletions, I is the number of insertions, C is the

number of correct words, N is the number of words in

the reference (N=S+D+C)

Character Error Rate (CER) is a prevalent perfor-

mance metric for automatic speech recognition sys-

tems. Similar to Word Error Rate (WER), CER op er-

ates at the ch aracter level rather than the word level.

For detailed insights (Morris et al., 2004 ). The com-

putation of the Character Error Rate involves:

CER = (S +D+I)/N = (S +D+I)/(S + D+C) (2)

where S is the number of substitutions, D is the num-

ber of deletions, I is the number of insertions, C is

the number of cor rect character s, N is the number of

characters in the referenc e (N=S+D+C) .

4.2.2 Manual Evaluation

In Manua l E valuation, we used Mean Opinion Score

(MOS) (Guski, 1997), it is a su bjective evaluation

metric used to measure the perceived quality of gen-

erated speech or audio. Huma n participants rate

the samples on a scale; the average score indicate s

the system’s quality. MOS evaluations help identify

strengths and weaknesses in TTS and ASR systems

and guide improvements for better user satisfaction.

It is one aspect of a com prehensive evaluation a p-

proach that consid ers other metrics and user feedback.

We conducted an anonymous survey to evaluate the

output of our mod el. 108 participants were asked to

rate the audio samples on a scale from 1 to 5, where

higher scores indicated better quality. The survey in-

cluded 10 audio samples, and we calculated the Mean

Opinion Score (MOS) based on the participants’ rat-

ings. This approach en sured the conﬁdentiality of the

participant’s responses and provided valuable insights

into the perceived quality of the model’s output.

5 RESULTS

The evaluation metrics of Character Error Rate

(CER), Word Error Rate (WER) based on 1 and 2,

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

224

(a) Ground-truth (b) Tacotron2 (c) Tacotron1

Figure 5: Comparison of predicted Mel-spectrograms for input text (



àA¿



Ë@



®K



Q¢Ë@ éJ

A«







ºÓ) (He did not like the way

that was).

(a) Ground-truth (b) Tacotron2 (c) Tacotron1

Figure 6: Comparison of predicted Mel-spectrograms for input text (@Yg





®¢



JÓ È@



ñ È



A



 A







'



èXð) (This leaves us to ask

a very logical question).

and Mean Opinion Score (MOS) are given in Table 4

and Table 5, Our best-performin g speech synthesizer

is Tacotron2 with HiFi-GAN for Egyptian Arabic.

The combination of the mel-spectrogram predic-

tion model of Tacotron2 and HiFi-GAN ha s demon-

strated supe rior performance in low CER and WER,

indicating accurate and precise speech generation.

Additionally, the high MOS scores obtained from

subjective evaluatio ns reﬂect the perceived quality

and naturalness of the synthesized speech. The re-

sults from these metrics collectively support the con-

clusion that Tacotron2 with HiFi-GAN is the most ef-

fective a nd preferred choice among the tested spe ech

synthesis mo dels for Egyptian Arabic. Furthermore ,

comparing this model’s (MOS) scores against those of

Modern Standard Arabic T TS sy stems p rovides addi-

tional validation of its superiority. This afﬁrmation

is underscored not only by the observed (Word Er-

ror Rate) (WER) and (Ch aracter Error Rate) (CER)

evaluations but also by the proximity and sur passing

of our achieved MOS scor es in relation to other sys-

tems in th e MSA context. Noteworthy examples in-

clude the NatiQ system (Abdelali et al., 2022), w here

our model excels, and the Transfer Learning End-

to-End Arabic Text-To-Sp e ech (TTS) Deep Ar c hitec-

ture (Fahmy et al., 2020), which further accentuates

the efﬁcacy of our approach. Additionally, in terms

of (WER) and (CER), our model surpasses the per-

formance of the NatiQ sy stem (Abdelali et al., 2022).

In Fig.5 and Fig.6, a comparison of Mel-

spectrograms for three different audios are presented

with two different sentences: a) the ground-truth,

b) Tacotron2, and c) Tacotron 1. Upon observation,

we can n otice that the re is a higher degree of simi-

larity between the ground-truth a nd Tacotron2 Mel-

spectrograms compared to Tacotron1. This ﬁnding

suggests that Tacotron2 is better at generating Mel-

spectrograms that closely resemble the gro und-truth,

indicating higher accuracy and ﬁdelity in the spee ch

synthesis p rocess. On the other hand, Tacotron1’s

Mel-spectrograms show more noticeable differences

from the ground-truth, suggesting that it may not cap-

ture speciﬁc acoustic characteristics as effectively as

Tacotron2.

Table 4: CER and WER evaluation results.

CER WER

NatiQ TTS System for (MSA) (Abdelali et al., 2022) 8.01 24.87

Tacotron1 (Griffen-Lim) for Egyptian Arabic 20.4 66.6

Tacotron2 (HiFi-GAN) for Egyptian Arabic (proposed) 7.3 22.3

Table 5: MOS evaluation results (maximum score is 5).

MOS

TL TTS System for (MSA) (Fahmy et al., 2020) 4.21

NatiQ TTS System for (MSA) (Abd elali et al., 2 022) 4.40

Tacotron1 (Griffen-Lim) for Egyptian Arabic 3.64

Tacotron2 (HiFi-GAN) for Egyptian Arab ic (propo sed) 4.48

Masry: A Text-to-Speech System for the Egyptian Arabic

225

6 CONCLUSION

This paper presents the Masry end-to-end text-to-

speech system tailored for Egyptian Arabic, combin-

ing Tacotron 2 with the HiFi-GAN Vocoder. Addition-

ally, a novel data set and its transcriptions in Egyptian

Arabic were introduced. T he system’s performance

was assessed through automatic evaluation metric s,

namely Charac te r Error Rate (CER) and Word Error

Rate (WER), resu lting in scores of 7 .3 and 22.3, re-

spectively. Furthermore , a manual evaluation using

the Mean Opinion Score (MOS) yielded a score of

4.48. Our ﬁndings indicate that the system’s perfor-

mance is in close proximity to that of English and

Modern Arabic Standard systems. Future work en-

tails incorporating additional features, such as emo-

tions and multispeaker capabilities, to enhance the

system’s cap abilities further.

REFERENCES

Abdel-Hamid, O., Abdou, S. M., and Rashwan, M. (2006).

Improving arabic hmm based speech synthesis qual-

ity. In Ninth International Conference on Spoken Lan-

guage Processing. n/a.

Abdel-Massih, E. T. (2011). An Introduction to Egyptian

Arabic. MPublishing, University of Michigan Library.

Abdelali, A., Durrani, N., Demiroglu, C ., Dalvi, F.,

Mubarak, H. , and Darwish, K. (2022). Natiq: An end-

to-end text-to-speech system for arabic. arXiv preprint

arXiv:2206.07373.

Alyafeai, Z. (2022). Klaam asr. https:// git hub.com/

ARBML/klaam.

Baali, M., Hayashi, T., Mubarak, H., Maiti, S., Watanabe,

S., El-Hajj, W., and Ali, A. (2023). Unsupervised data

selection for TTS: using arabic broadcast news as a

case study. CoRR, abs/2301.09099.

El-Imam, Y. A. (2004). Phonetization of arabic: rules

and algorithms. Computer Speech & Language,

18(4):339–373.

Fahmy, F. K., Khalil, M. I., and Abbas, H. M. (2020). A

transfer learning end-to-end arabic text-to-speech (tts)

deep architecture. In IAPR Workshop on Artiﬁcial

Neural Networks in Pattern Recognition, pages 266–

277. Springer.

Guski, R. (1997). P sychological methods for evaluat-

ing sound quality and assessing acoustic information.

Acta Acustica united with Acustica, 83:765–774.

Habash, N. Y. (2022). Introduction to Arabic natural lan-

guage processing. Springer N ature.

Halabi, N. (2016). Modern standard Arabic phonetics for

speech synthesis. PhD thesis, University of Southamp-

ton.

Imene, Z., Mnasri, Z., Vincent, C., D enis, J., Amal, H.,

et al. (2018). Duration modeling using dnn f or arabic

speech synthesis. In Proeedings of 9th International

Conference on Speech Prosody, pages 597–601.

Ito, K. and Johnson, L. (2017). The lj speech dataset. https:

//keithito.com/LJ-Speech-Dataset/.

Kong, J., K im, J., and Bae, J. (2020). Hiﬁ-gan: Genera-

tive adversarial networks for efﬁcient and high ﬁdelity

speech synthesis. Advances in Neural Information

Processing Systems, 33:17022–17033.

Morris, A., Maier, V., and Green, P. (2004). From wer

and r il to mer and wil: improved evaluation measures

for connected speech recognition. In INTERSPEECH

2004 - ICSLP, 8th International Conference on Spoken

Language Processing, Jeju Island, Korea.

Obeid, O., Zalmout, N., Khalifa, S., Taji, D. , Oudah, M.,

Alhafni, B., Inoue, G., Eryani, F., Erdmann, A., and

Habash, N. (2020). CAMeL tools: An open source

python toolkit for Arabic natural language processing.

In Proceedings of the Twelfth Language Resources

and Evaluation Conference, pages 7022–7032, Mar-

seille, France. European Language Resources Associ-

ation.

Ren, Y., Ruan, Y., Tan, X., Qin, T., Zhao, S., Zhao, Z., and

Liu, T.-Y. (2019). Fastspeech: Fast, robust and con-

trollable text to speech. Advances in neural informa-

tion processing systems, 32.

Shen, J., Pang, R., Weiss, R. J., Schuster, M., Jaitly, N.,

Yang, Z., Chen, Z., Zhang, Y., Wang, Y., Skerrv-Ryan,

R., et al. (2018). Natural tts synthesis by condition-

ing wavenet on mel spectrogram predictions. In 2018

IEEE international conference on acoustics, speech

and signal processing (ICA SSP), pages 4779–4783.

IEEE.

Versteegh, K. (2014). Arabic language. Edinburgh Univer-

sity Press.

Wang, Y., Skerry-Ryan, R. J., Stanton, D., Wu, Y., Weiss,

R. J., Jaitly, N., Yang, Z., Xiao, Y., Chen, Z., Bengio,

S., Le, Q. V., Agiomyrgiannakis, Y., Clark, R. A. J.,

and Saurous, R. A. (2017). Tacotron: Towards end-

to-end speech synthesis. In Interspeech.

Young, M., Courtad, C. A., Douglas, K., and Chung, Y.-

C. (2018). The effects of text-to-speech on reading

outcomes for secondary st udents with learning dis-

abilities. Journal of Special Education Technology,

34:016264341878604.

ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics

226