This is a TTS model pretrained on the pre-tokenized Emilia dataset. Since there's no speaker conditioning, the speaker is random at inference. This model has 550M parameters and it was trained from scratch on a single AMD Instinct MI300X for ~4 days with the ROCm PyTorch Training v25.7 container.
The training objective was standard next-token prediction on concatenated text-audio tokens.
Code
For more information including a Colab notebook, see the repository.
