NileTTS: Egyptian Arabic Text-to-Speech Dataset 🇪🇬

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

NileTTS is the first large-scale, publicly available Egyptian Arabic (اللهجة المصرية) text-to-speech dataset, comprising 38 hours of transcribed speech across diverse domains.

Dataset Description

• Total Duration: 38.1 hours
• Total Utterances: 9,521
• Speakers: 2 (1 male, 1 female)
• Domains: Medical, Sales & Customer Service, General Conversations
• Language: Egyptian Arabic (colloquial)
• Audio Format: WAV, 24kHz
• Average Utterance Length: 14.4 seconds

Dataset Statistics

Data Collection Pipeline

NileTTS was constructed using a novel synthetic data generation pipeline:

1. Content Generation

Large language models (Gemini, Claude) generate Egyptian Arabic content across three domains:

• Medical consultations and health discussions
• Sales and customer service scenarios
• General everyday conversations

2. Audio Synthesis

The generated content is converted to natural Egyptian Arabic speech using NotebookLM's audio generation capabilities, which produces high-quality podcast-style audio with two distinct speakers.

3. Transcription & Segmentation

• Audio is transcribed using Whisper Large with Arabic language setting
• Long audio files are segmented into utterances
• Timestamps are preserved for alignment

4. Speaker Diarization

• Speaker embeddings extracted using ECAPA-TDNN (SpeechBrain)
• K-Means clustering identifies the two speaker centroids from initial samples
• Each utterance is assigned to the nearest speaker centroid

5. Quality Control

• Manual verification of transcriptions for accuracy
• Speaker label validation
• Removal of low-quality or misaligned segments

Metadata Format

The metadata CSV files follow the XTTS training format:

audio_file|text|speaker_name
wav/sales_audioid_chunkidx.wav|مرحبا، إزيك النهارده؟|SPEAKER_01
wav/medical_audioid_chunkidx.wav|أهلا وسهلا، الحمد لله تمام|SPEAKER_02
wav/general_audioid_chunkidx.wav|أهلا وسهلا، الحمد لله تمام|SPEAKER_02
...

Usage

Interactive Demo

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("KickItLikeShika/NileTTS")

Training TTS Models

The dataset is formatted for direct use with XTTS v2 fine-tuning:

import pandas as pd

train_df = pd.read_csv("metadata_train.csv", sep="|")
eval_df = pd.read_csv("metadata_eval.csv", sep="|")

Intended Use

• Primary: Training and fine-tuning text-to-speech models for Egyptian Arabic
• Secondary:
  • Egyptian Arabic ASR training/evaluation
  • Speaker recognition research
  • Dialectal Arabic NLP research

Limitations

• Speaker Diversity: Limited to 2 speakers (1 male, 1 female)
• Synthetic Origin: Audio was generated synthetically, not recorded from human speakers
• Domain Coverage: Focused on three specific domains; may not generalize to all topics
• Dialect Variation: Represents one style of Egyptian Arabic; regional variations within Egypt are not captured

Citation

If you use this dataset, 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"
}

Related Resources

• Fine-tuned Model: KickItLikeShika/Nile-XTTS-v2
• Base Model: coqui/XTTS-v2

License

This dataset is released under the Apache 2.0 license.