CoSTA decouples speech token spaces for efficient dialogue models
October 14, 2026
CoSTA is a post-hoc framework that decouples generation and synthesis token spaces in Spoken Dialogue Models. By using mean-initialized vocabulary extension and pronunciation adaptation, it enables the use of lower-rate speech tokenizers to reduce inference costs.
HOW THIS AFFECTS YOU
●
builderYou can significantly reduce latency and inference costs for voice-based AI agents.
●
researcherThis method allows for efficient token compression without breaking the coupling of generation and synthesis.