🗣️ Natq – Arabic Text-to-Speech (TTS) System

📌 Overview

Natq is an open-source Arabic Text-to-Speech (TTS) system that uses deep learning models to convert Arabic text into natural-sounding speech.
We designed and trained multiple architectures on public datasets with a focus on Arabic language intricacies, including diacritization and phoneme modeling.

---

Graduation Project – Faculty of Computing and Artificial Intelligence, Helwan University
Department of Artificial Intelligence
Supervised by: Dr. Yasser Hifny & Dr. Ahmed Hesham

🎯 Objectives

• Create a modular and efficient Arabic TTS system using public datasets.
• Handle Arabic-specific challenges such as diacritics and phoneme mapping.
• Compare performance of different spectrogram generators and vocoders.
• Support live speech synthesis via a web app (React + FastAPI).

---

🧠 Architecture

The system pipeline is composed of:

1. Text Preprocessing
2. Diacritization: CATT Transformer
3. Grapheme-to-Phoneme (G2P): Nawar Halabi’s Phenomizer
4. Spectrogram Generation: FastPitch, FastSpeech2, Mixer-TTS, or Spark-TTS
5. Vocoder: HiFi-GAN for waveform synthesis

---

📚 Datasets

Dataset	Hours	Diacritized	Accent	Notes
Arabic_Speech_Corpus	4.1	✅	Levantine	Open-source
ClArTTS	12	✅	Classical MSA	Based on LibriVox audiobook

---

🛠 Models Used

🔹 Spectrogram Generators

• FastPitch – Parallel and pitch-controllable TTS
• FastSpeech2 – Variance-aware and efficient
• Mixer-TTS – MLP-Mixer based parallel synthesis
• Spark-TTS – End-to-end LLM-based TTS with zero-shot speaker cloning

🔹 Vocoder

• HiFi-GAN – Fast, high-fidelity waveform generation

---

🧪 Evaluation

Objective

• Spectrogram quality
• Inference time
• Model size

Subjective

• Mean Opinion Score (MOS) from human listeners

---

🔈 Audio Samples

وَالسَّلَامُ عَلَى أَشْرَفِ الْأَنْبِيَاءِ وَالْمُرْسَلِينَ سَيِّدِنَا مُحَمَّدٍ

Model	Dataset	Sample
FastPitch	ClArTTS	▶️ Listen
Mixer-TTS	ASC	▶️ Listen
Spark-TTS	ClArTTS	▶️ Listen
VITS_facebook	—	▶️ Listen
T5_MBZUAI	—	▶️ Listen
FastSpeech2	—	▶️ Listen

---

👥 Contributors

• Abdelhalim Ashraf
• Abdelrahman Ramadan
• Eid Osama Eid
• Omar Tamer
• Ali Adel

---

📂 Project Structure

d:/coding/Natq/
├── .gitignore
├── FastPitch/
│   ├── .vscode/
│   │   └── sttings.json
│   ├── arabic_phoneme_tokenizer.py
│   ├── catt/
│   ├── custom_arabic_to_phones.py
│   └── ...
├── FastSpeech2/
│   ├── arabic_phoneme_tokenizer.py
│   ├── catt/
│   └── ...
├── Media/
│   ├── audio_samples/
│   │   ├── FastPitch-TTS_MOS/
│   │   │   ├── bism_fp.mp4
│   │   │   ├── bism_fp.wav
│   │   │   └── ...
│   │   ├── Mixer-TTS_MOS/
│   │   │   ├── bism_mix.mp4
│   │   │   ├── bism_mix.wav
│   │   │   └── ...
│   │   └── Spark-TTS_MOS/
│   │       ├── bism_spark.mp4
│   │       ├── bism_spark.wav
│   │       └── ...
│   └── images/
│       ├── Demo.jpg
│       └── Fully_Logo_White.png
├── Natq-Frontend/
│   ├── .gitignore
│   ├── eslint.config.js
│   └── ...
├── README.md
└── Spark/
    ├── catt/
    ├── download_model_files.ipynb
    └── ...

---

🌐 Web Application

We built a simple TTS web interface using FastAPI for the backend and React for the frontend.

🧪 Test it locally:

cd backend
uvicorn main:app --reload

Then in another terminal

cd ../app
npm install
npm start