This paper describes a general and effective domain selection framework for multi-domain spoken dialogue systems that employ distributed domain experts. The framework consists of two processes: deciding if the current domain continues and estimating the probabilities for selecting other domains. If the current domain does not continue, the domain with the highest activation probability is selected. Since those processes for each domain expert can be designed independently from other experts and can use a large variety of information, the framework achieves both extensibility and robustness against speech recognition errors. The results of an experiment using a corpus of dialogues between humans and a multi-domain dialogue system demonstrate the viability of the proposed framework.