CoSTA framework for efficient spoken dialogue model inference

CoSTA decouples generation and synthesis token spaces in spoken dialogue models to improve efficiency. The framework achieves 1.7x faster speech-token generation and lower latency by using a compressed representation for the generation phase.