TTS foundation model compatible with Llama framework (160k hours tokenized speech data released)