Llasa

Llasa frames text-to-speech as ordinary language modeling: speech is quantized into a single XCodec2 codebook of 65,536 tokens, and a transformer initialized from Llama at 1B, 3B, and 8B scales learns to continue text sequences with speech tokens, trained on 250,000 hours of Chinese and English audio. Because the interface is just tokens, the models support zero-shot voice cloning from a short reference clip and expressive synthesis driven by textual context, and the accompanying paper (arXiv 2502.04128) studies how both train-time and inference-time compute scaling improve naturalness and prosody. Inference runs through Hugging Face transformers with the XCodec2 package on a CUDA GPU, and the HKUST audio lab publishes the checkpoints on Hugging Face alongside a separate training repository, LLaSA_training, for reproducing or fine-tuning the models. The weights are licensed CC-BY-NC-4.0, restricting them to non-commercial use, which keeps Llasa a research and prototyping option rather than a drop-in commercial TTS engine, though the family remains popular for voice cloning experiments.