CORTIS: Text-Only Adaptation of Spoken Language Models for Task-Oriented Voice Agents

CORTIS is introduced as a text-only adaptation framework for task-oriented voice agents, allowing fine-tuning of spoken language models (SLMs) using text-form task supervision without the need for paired speech-target annotations. Evaluated on two Qwen2.5-Omni backbones across three task-oriented speech datasets, CORTIS demonstrates competitive performance with ASR-LLM cascades, particularly excelling in preserving high-level task semantics under acoustic degradation. This approach offers a practical solution for practitioners facing challenges in collecting speech-target data, enhancing the robustness of voice agents in real-world applications.