Alania-2 is a 488M-parameter Turkish text-to-speech model using a VoxCPM architecture with a diffusion head. Trained on 18,195 hours of speech, it achieves a 1.44% word error rate on the Freya-TR-Eval benchmark, significantly improving upon its predecessor's 5.99%.
HOW THIS AFFECTS YOU
●
builderYou can use this for high-accuracy Turkish voice synthesis and voice cloning.
●
researcherThe model uses a language-model backbone with a diffusion head trained largely on synthetic data.