Audar-TTS-V1
A family of three Arabic-first, zero-shot text-to-speech models that clone a voice from a short reference clip, support inline expression tags, and speak Modern Standard and several dialectal forms of Arabic as well as English.
Audar-TTS-V1 is a family of three text-to-speech models - Flash (0.55B), Turbo (1.64B) and Pro (4B) - built by AudarAI around Arabic speech. Instead of a phonemizer or per-language pronunciation rules, synthesis is framed as next-token prediction over a joint vocabulary of text, control tokens and discrete speech codes, so dialect coverage comes from training data. Each model clones a voice from a 5 to 15 second reference clip with no per-speaker fine-tuning, and shapes delivery with inline expression tags such as laughs, whispers or excited.
The models speak Modern Standard Arabic, several regional dialects, code-switched Arabic-English, and English, at 24 kHz, and were trained on over 200,000 hours of audio-text data with a preference-alignment stage scored on naturalness, similarity, prosody, expression fidelity and pronunciation. On public Arabic benchmarks judged by automatic speech recognition, the Pro tier scores a statistically lower word error rate than GPT-4o-mini-TTS.
Flash and Turbo are downloadable in Transformers and GGUF formats; Pro is available only through AudarAI's hosted API.