PatientDesk AI Releases 30MB Antalia-2 Mini Speech Model

PatientDesk AI has released Antalia-2 Mini, an open-source Turkish text-to-speech model licensed under Apache-2.0. Occupying just 30.5 MB in fp32 with 7.62 million parameters, the ultra-compact system rivals speech generation tools over 300x larger. The model leverages flow matching paired with shortcut self-consistency, enabling developers to generate speech in 1, 2, 4, or 8 inference steps. The development team trained the system using 1,578 hours of synthetic audio generated by VoxCPM2, completely omitting human speech recordings.

In performance testing on the Freya-TR-Eval benchmark across 15 Turkish speech implementations, Antalia-2 Mini secured the lowest character error rate (CER) of 0.19%. On word error rate (WER), it effectively tied with EMA Lightning, placing just 0.01 percentage points behind—a distinction within normal seed-to-seed variance. It achieved lower CER scores than significantly larger systems, including the 2.38 billion-parameter Trendyol-TTS, the 500 million-parameter Chatterbox Multilingual, the 336 million-parameter Anka TTS, and Gemini 3.8 Flash TTS.

The model is optimized for CPU-only local execution, requiring no dedicated GPU. On an Apple M5 CPU operating on 2 threads, Antalia-2 Mini synthesizes speech at speeds 50x faster than real time. PatientDesk AI has published the model weights, text normalizer, and training recipe. Practitioners can test the technology via a Hugging Face Spaces browser demo or install the library directly through pip install antalia-mini.

For developers building Turkish-language applications, Antalia-2 Mini eliminates the standard dilemma of choosing between expensive cloud APIs, heavy multi-gigabyte models, or inaccurate lightweight engines. By concentrating on a single voice in one language, the system offers high accuracy and minimal latency on consumer hardware, making edge speech synthesis practical for resource-constrained environments.