NileTTS

Paper: https://arxiv.org/abs/2602.15675

NileTTS is a large-scale Egyptian Arabic Text-to-Speech dataset and fine-tuned XTTS model. This repository contains the code for data generation, model training, and evaluation as described in our paper.

Resources

Overview

NileTTS addresses the lack of high-quality TTS resources for Egyptian Arabic by providing:

1. 38 hours of transcribed Egyptian Arabic speech across medical, sales, and general conversation domains
2. A fine-tuned XTTS v2 model optimized for Egyptian Arabic synthesis
3. A reproducible synthetic data generation pipeline

Results

Quick Start

Installation

pip install -r requirements.txt

Using the Model

See playground.ipynb for a complete example of loading and using the model and the dataset.

Repository Structure

NileTTS/
├── generate-data.py
├── evaluate.py
├── playground.ipynb
├── requirements.txt
└── README.md

Data Generation Pipeline

The generate-data.py script processes audio files generated by NotebookLM into training-ready chunks with transcriptions and speaker labels.

Prerequisites

Before running the script, you need:

1. Audio file: An .m4a or .wav file containing Egyptian Arabic speech (e.g., from NotebookLM)
2. Speaker centroids: A speaker_centroids.pkl file containing pre-computed speaker embeddings for diarization

How It Works

1. Transcription: Uses Whisper Large to transcribe the audio with Arabic language setting
2. Chunking: Groups transcription segments into chunks of max 15 seconds
3. Speaker Diarization: Identifies speaker for each chunk using ECAPA-TDNN embeddings and cosine similarity to pre-computed centroids
4. Export: Saves audio chunks as WAV files with corresponding transcriptions and metadata CSV

Interactive Demo

The playground.ipynb notebook demonstrates:

• Loading the NileTTS model from HuggingFace
• Downloading dataset samples on-demand
• Generating audios/speech
• Playing and saving generated audio

Citation

If you use NileTTS in your research, please cite:

@inproceedings{khamis-ahmed-2026-llm,
    title = "{LLM}-to-Speech: A Synthetic Data Pipeline for Training Dialectal Text-to-Speech Models",
    author = "Khamis, Ahmed  and
      Ahmed, Hesham Ali",
    booktitle = "Proceedings of the 2nd Workshop on {NLP} for Languages Using {A}rabic Script",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.abjadnlp-1.6/",
    pages = "47--54"
}

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

Acknowledgements

• XTTSv2-Finetuning-for-New-Languages by @anhnh2002 for the training code. We adapted their finetuning pipeline and implemented additional evaluation metrics (WER, CER, Speaker Similarity) and Weights & Biases integration.
• Coqui TTS for the XTTS v2 architecture
• OpenAI Whisper for transcription
• SpeechBrain for speaker embeddings
• Google NotebookLM for audio synthesis