https://github.com/barqawiz/shakkala
Deep learning for AR text Vocalization - التشكيل الالي للنصوص العربية
https://github.com/barqawiz/shakkala
arabic-nlp deep-learning lstm rnn
Last synced: 2 months ago
JSON representation
Deep learning for AR text Vocalization - التشكيل الالي للنصوص العربية
- Host: GitHub
- URL: https://github.com/barqawiz/shakkala
- Owner: Barqawiz
- License: other
- Created: 2017-11-28T04:48:14.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2023-03-25T01:50:27.000Z (about 2 years ago)
- Last Synced: 2024-10-18T00:44:33.413Z (8 months ago)
- Topics: arabic-nlp, deep-learning, lstm, rnn
- Language: Python
- Homepage: https://shakkala.com
- Size: 578 MB
- Stars: 337
- Watchers: 25
- Forks: 44
- Open Issues: 5
-
Metadata Files:
- Readme: README.md
- Funding: .github/FUNDING.yml
- License: LICENSE.md
Awesome Lists containing this project
README
# Shakkala Project مشروع شكّالة
[](https://badge.fury.io/py/shakkala)
## Introduction
The Shakkala project presents a recurrent neural network for Arabic text vocalization that automatically forms Arabic characters (تشكيل الحروف) to enhance text-to-speech systems. The model can also be used in other applications such as improving search results. In the beta version, the model was trained on over a million sentences, including a majority of historical Arabic data from books and some modern data from the internet. The accuracy of the model reached up to 95%, and in some data sets it achieved even higher levels of accuracy depending on complexity and data distribution. This innovative approach has the potential to significantly improve the quality of writing and text-to-speech systems for the Arabic language.
## Requirements
### Easy setup
No GitHub repository installation is needed for [pip](https://pypi.org/project/shakkala/) case:
```
pip install shakkala
```
### Project setup
Execute the source code from Github:
```
cd requirements
pip install -r requirements.txt
cd ..
```
Note: Shakkala has been tested with Tensorflow 2.9.3.
## Code Examples (How to)
Check full example in (demo.py) file.
0. Import
```
from shakkala import Shakkala
```
1. Create Shakkala object
```
sh = Shakkala()
```
OR for advanced usage:
```
sh = Shakkala(version={version_num})
```
2. Prepare input
```
input_text = "فإن لم يكونا كذلك أتى بما يقتضيه الحال وهذا أولى"
input_int = sh.prepare_input(input_text)
```
3. Call the neural network
```
model, graph = sh.get_model()
logits = model.predict(input_int)[0]
```
4. Predict output
```
predicted_harakat = sh.logits_to_text(logits)
final_output = sh.get_final_text(input_text, predicted_harakat)
```
Available models:
- version_num=1: First test of the solution.
- version_num=2: Main release version.
- version_num=3: Some enhancements from version number 2.
It worth to try both version_num=2 and version_num=3.
### Demo run
The fastest way to start with Shakkala by running the demo from Github:
```
python demo.py
```
## Perfomance Tips
Shakkala built in object oriented way to load the model once into memory for faster prediction, to make sure you dont load it multiple times in your service or application follow the steps:
- Load the model in global variable:
```
sh = Shakkala(folder_location, version={version_num})
model, graph = sh.get_model()
```
- Then inside your request function or loop add:
```
input_int = sh.prepare_input(input_text)
logits = model.predict(input_int)[0]
predicted_harakat = sh.logits_to_text(logits)
final_output = sh.get_final_text(input_text, predicted_harakat)
```
## Accuracy
In this beta version 2 accuracy reached up to 95% and in some data it reach more based on complexity and data disribution.
This beta version trained on more than million sentences with majority of historical Arabic data from books and **some of** available formed modern data in the internet.
### Prediction Example
For live demo based on Shakkala library click the [link](http://ahmadai.com/shakkala/)
| Real output | Predicted output |
| ------------- | ---------------- |
| فَإِنْ لَمْ يَكُونَا كَذَلِكَ أَتَى بِمَا يَقْتَضِيهِ الْحَالُ وَهَذَا أَوْلَى | فَإِنْ لَمْ يَكُونَا كَذَلِكَ أَتَى بِمَا يَقْتَضِيهِ الْحَالُ وَهَذَا أَوْلَى |
| قَالَ الْإِسْنَوِيُّ وَسَوَاءٌ فِيمَا قَالُوهُ مَاتَ فِي حَيَاةِ أَبَوَيْهِ أَمْ لَا | قَالَ الْإِسْنَوِيُّ وَسَوَاءٌ فِيمَا قَالُوهُ مَاتَ فِي حَيَاةِ أَبَوَيْهِ أَمْ لَا |
| طَابِعَةٌ ثُلَاثِيَّةُ الْأَبْعَاد | طَابِعَةٌ ثَلَاثِيَّةُ الْأَبْعَادِ |
### Accuracy Enhancements
The model can be enhanced to reach more than 95% accuracy with following:
- Availability of more formed **modern** data to train the network. (because current version trained with mostly available historical Arabic data and some modern data)
- Stack different models
## Model Design
## References
- A paper compare different arabic text diacritization models and show that shakkala is the best among available neural networks for this solution:
[Arabic Text Diacritization Using Deep Neural Networks, 2019](https://arxiv.org/abs/1905.01965)
## Citation
For academic work use
```
Shakkala, Arabic text vocalization, Barqawi & Zerrouki
```
OR bibtex format
```
@misc{
title={Shakkala, Arabic text vocalization},
author={Barqawi, Zerrouki},
url={https://github.com/Barqawiz/Shakkala},
year={2017}
}
```
## Contribution
### Core Team
1. Ahmad Barqawi: Neural Network Developer.
2. Taha Zerrouki: Mentor Data and Results.
### Contributors
1. Zaid Farekh & propellerinc.me: Provide infrastructure and consultation support.
2. Mohammad Issam Aklik: Artist.
3. Brahim Sidi: Form new sentences.
4. Fadi Bakoura: Aggregate online content.
5. Ola Ghanem: Testing.
6. Ali Hamdi Ali Fadel: Contribute code.
License
-------
Free to use and distribute only mention the original project name Shakkala as base model.
The MIT License (MIT)
Copyright (c) 2017 Shakkala Project
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.