Turkish TTS model ditches tokenizers for direct latent diffusion

FreyaTTS operates end-to-end from raw characters without phonemizers or discrete tokenization, which matters because tokenizer design has been a hidden failure point for non-Latin scripts. The related Bengali ASR work showed that English-centric tokenization schemes fragment morphologically complex languages into excessive token chains, destabilizing inference. FreyaTTS sidesteps this problem entirely by learning the character-to-audio mapping directly.

This connects directly to the tokenizer transplantation work from earlier this month. That paper exposed how efficiency gains on Latin-script benchmarks often come at the cost of catastrophic failure on underrepresented languages. FreyaTTS takes the opposite approach: rather than fixing tokenizers after the fact, it eliminates the intermediate representation layer altogether. The same design principle appears in the CoCoT-EEG work, which rejected masked pretraining plus tokenization for biomedical signals in favor of direct contrastive learning on raw data. Both papers suggest a pattern where forcing diverse modalities through standardized tokenization schemes is a liability, not a feature.

If FreyaTTS achieves comparable quality on morphologically rich languages (Arabic, Finnish, Hungarian) without retraining the core architecture, that confirms the tokenizer-free approach generalizes. If the model requires language-specific tuning of the diffusion parameters or VAE, the advantage over tokenizer transplantation narrows significantly.