The AuK technical report introduces an open-source foundational model that unifies speech generation and editing through natural-language instructions and audio context.

  • The model is trained on approximately 3.03 billion instruction-audio instances and 1.95 million hours of supervision across five task families.
  • Architecture combines a multimodal large language model, a VAE for acoustic conditioning, and a hybrid rectified-flow Transformer.
  • Post-training includes human-feedback preference optimization and reward-based reinforcement learning.
  • AuK-Flash achieves 4-step inference with a 4.5 wall-clock speedup over the full model.

The authors release source code and model weights to support reproducibility and further research.