CoSTA-EN CoSTA: Compressed Speech Token Adaptation for Efficient Spoken Dialogue Models https://arxiv.org/abs/2610.21143 https://github.com/yingzhuodeng/CoSTA Based on Step-Audio-2-mini by StepFun: https://huggingface.co/stepfun-ai/Step-Audio-2-mini The upstream model and implementation are licensed under Apache-2.0. CoSTA adds compressed speech-token vocabulary adaptation and dialogue fine-tuning. The custom model code includes changes for training loss, embedding/LM-head resizing, and inference timing. Existing source notices are retained.