Democratizing Arabic AI by centralizing the world's best open-source LLMs, Speech models, and Datasets.
Awesome-Arabic-AI is the definitive hub for advancements in Arabic Large Language Models (LLMs), Text-to-Speech (TTS), and Speech-to-Text (STT) technologies. Arabic, spoken by over 400 million people, presents unique computational challengesโfrom its complex morphology and lack of diacritics to the profound state of diglossia (the gap between MSA and spoken dialects).
Our mission is to centralize high-quality, open-weight models and research benchmarks to provide a strategic foundation for developers navigating the burgeoning Arabic AI ecosystem. We aim to foster a community where state-of-the-art speech and language technologies are accessible and optimized for the linguistic richness of the Arab world.
The latest and most significant breakthroughs in the Arabic AI space:
- ๐ฃ๏ธ SILMA TTS v1: A lightweight (150M) bilingual Arabic-English model with ultra-low latency and instant voice cloning. Demo
- ๐ฃ๏ธ Lahgtna: A state-of-the-art Arabic Dialect TTS model supporting Egyptian, Saudi, Moroccan, and Iraqi dialects with full diacritics support.
- ๐ง Karnak LLM: A 40B depth-extended Arabic-English model optimized for reasoning and long-context (20K tokens). It is built on top of Qwen3-30B-A3B.
- ๐ ๏ธ YouTube Audio Extractor: A powerful tool for building Egyptian ASR datasets from YouTube, featuring AI transcription (Gemini 2.0 Flash) and VAD filtering.
- ๐ฌ Research Spotlight: NAMAA Space
- ๐๏ธ Arabic TTS Excellence
- ๐ฐ Commercial Arabic Voice
- ๐๏ธ Speech-to-Text (STT)
- ๐ง Large Language Models (LLMs)
- ๐ ๏ธ Tools & Utilities
- ๐ Dialectal AI & Community Hubs
- ๐ Datasets & Benchmarks
- ๐ Research & Publications
- ๐ฅ Key Researchers & Organizations
- ๐ค Contributing
Network for Advancing Modern ArabicNLP & AI
| Model / Project | Type | Description | Link |
|---|---|---|---|
| AraModernBERT | LLM (Encoder) | SOTA Arabic ModernBERT (Base V1.0), optimized for 8K context length. | HuggingFace |
| Qari-OCR v0.3 | Vision/OCR | High-fidelity Arabic OCR based on Qwen2-VL, supports diacritics & complex layouts. | HuggingFace |
| NAMAA-MT-Saudi | Translation | Saudi Dialect (Najdi, Hijazi) to English translation system. | HuggingFace |
| Namaa-Reranker | RAG Tool | High-performance reranker fine-tuned on mMARCO for Arabic IR. | HuggingFace |
| SaudiSpell-AraT5 | Spelling Correction | SOTA sequence-to-sequence spelling correction for Saudi dialects (Najdi, Hijazi) and MSA. | HuggingFace |
These models represent the cutting edge of open-source Arabic speech synthesis.
| Model Name | Developer/Author | Links | Notes |
|---|---|---|---|
| SILMA TTS v1 | Silma AI | Model / Demo | Lightweight (150M) bilingual (AR/EN) model with instant voice cloning |
| Lahgtna | oddadmix | Space / Model | Multi-dialect TTS (EG, SA, MA, IQ) with Tashkeel support |
| Arabic-F5-TTS-v2 | Ibrahim Salah | Model | Advanced F5-based Arabic TTS |
| Arabic-TTS-Spark | Ibrahim Salah | Space | Spark-based Arabic synthesis |
| Fish Speech S2 Pro | Fish Audio | Model | SOTA multi-lingual speech model |
| MOSS-TTS | OpenMOSS-Team | Model | Open-source multi-lingual TTS |
| KaniTTS Arabic | nineninesix | Model | 400M parameter Arabic specialized TTS |
| Multilingual Chatterbox | Resemble AI | Model | Resemble AI multilingual foundation model |
| OuteTTS 1.0 | OuteAI | Model | Llama-based 1B parameter TTS |
| Fish Speech S1-mini | Fish Audio | Model | Lightweight high-quality speech model |
| Habibi-TTS | SWivid | Model/Dataset | High-quality Arabic speech dataset & model |
| SpeechT5 Arabic | MBZUAI | Model | MBZUAI fine-tuned SpeechT5 for Arabic |
| XTTS-v2 (Arabic) | Coqui/Community | Link | Multi-lingual support with Arabic fine-tuning |
Premium Arabic voice synthesis and AI services.
| Service Name | Developer | Links | Notes |
|---|---|---|---|
| Ziila | Intella | Link | Arabic native digital human for fluid customer experiences |
| Hamsa | Hamsa AI | Try it out | Commercial Arabic voice platform |
| SILMA TTS Voice | Silma AI | App | Provide a Commercial Arabic voice |
- ArTST v2: Unified Transformer for Arabic text and speech, supporting 17 dialects (MBZUAI).
- Whisper Large V3 Turbo (Arabic): High-speed throughput fine-tuned for Arabic.
- Whisper-Large-V2-Arabic-5k: High accuracy on Common Voice.
- Whisper-Medium-Egyptian: Specialized for Egyptian Arabic dialect.
- Falcon-H1-Arabic (3B, 7B, 34B): Latest models from TII with hybrid Mamba-Transformer architecture for superior Arabic performance (Jan 2026).
- Jais (G42): The leading Arabic-centric foundational model.
- Falcon 3 (TII): High-performance multilingual models.
- Allam (KSAA): Saudi model.
- Karnak (AIC): 40B depth-extended model built on Qwen3, featuring an Arabic-optimized tokenizer and 20K context window. It is built on top of Qwen3-30B-A3B.
- Arabic-llama3.1-16bit-FT: Fine-tuned on BigScience xP3.
- HeshamHaroon/Arabic-llama3: Fine-tuned from Meta Llama 3.
Specialized tools for Arabic data processing and model deployment.
- YouTube Audio Extractor: A comprehensive web application for building Egyptian Arabic ASR datasets. Supports channel monitoring, audio extraction via
pytubefix, and Google Gemini 2.0 Flash transcription. - Camel-tools: Essential suite for Arabic NLP (morphology, NER, sentiment).
- PyArabic: Library for Arabic text manipulation and normalization.
Arabic is not just MSA. This section collects resources for specific regional dialects.
- Lahgtna (Egyptian): Multi-dialect model with dedicated support for Egyptian Arabic.
- Chatterbox-Egyptian: State-of-the-art synthesis for Egyptian colloquialisms.
- EGTTS-v0.1: Community effort for Egyptian dialect speech.
- Lahgtna (Saudi): High-quality Saudi dialect synthesis with full diacritics.
- Saudi Dialect Hub (NAMAA): A unified hub for Saudi datasets and the SAFIR Leaderboard.
- NAMAA-MT-Saudi: Translation systems for Najdi and Hijazi.
- SaudiSpell-AraT5: SOTA sequence-to-sequence spelling correction for Saudi dialects.
- Silma TTS Benchmark: The standard for open-source Arabic TTS evaluation.
- OALL (Open Arabic LLM Leaderboard): The standard for LLM evaluation.
- Habibi Dataset: Critical high-quality audio dataset.
- Common Voice (Arabic): Massive community-driven speech dataset.
- SAFIR Leaderboard: Evaluation specifically for Saudi Arabic tasks.
- Baseer: Arabic Document OCR (Sep 2025): Vision-Language model for high-fidelity OCR, achieving SOTA 0.25 WER.
- NAMAA Space: Institutional-grade research in OCR, LLMs, and Dialects.
- oddadmix: Focus on Dialectal/Egyptian AI.
- Ibrahim Salah: Specialist in F5 and Spark TTS.
- SWivid: Creators of the Habibi dataset.
- Silma AI: Leading Arabic Benchmarks and high-end models.
- Intella: Developers of Ziila, focus on Arabic native digital humans and STT.
- TII (Technology Innovation Institute): Developers of Falcon and primary Arabic leaderboards.
- Applied Innovation Center (AIC): Developers of Karnak LLM, focused on depth-extended models for Arabic.
We welcome contributions! Please see CONTRIBUTING.md for guidelines. Researchers are encouraged to submit their models via Pull Request.
Distributed under the Apache 2.0 License. See LICENSE for more information.