Freya-TTS releases tokenizer-free Turkish TTS model with 8.0% WER

Researchers introduce Freya-TTS, a compact, tokenizer-free text-to-speech model optimized for Turkish that achieves an 8.0% word error rate on the Freya-TR-Eval benchmark. The 183.2M-parameter non-autoregressive conditional flow-matching Diffusion Transformer operates in the frozen continuous latent space of AudioVAE2, enabling high-quality 48 kHz reconstruction without a phonemizer or discrete speech tokenizer.

• Uses a 92-symbol Turkish character vocabulary for rule-free end-to-end modeling.
• Employs non-autoregressive parallel denoising to predict the entire latent sequence simultaneously.
• Utilizes a two-stage post-training recipe involving single-speaker voice locking and short-utterance coverage.
• Achieves a real-time factor of 0.11 on consumer GPUs and runs faster than real time on laptop CPUs.

The model is released under the Apache-2.0 license, offering efficient conversational synthesis suitable for resource-constrained edge deployment.