This work attempts to improve recognition accuracy by predicting the next utterance in a dialog. We propose a dialog-conditioned stochastic language model that is applied to dialog speech recognition. Each dialog-conditioned stochastic language model has been constructed with text data of a certain situation and implemented using a Japanese syllable trigram. Each situation was denned in advance to predict next utterances effectively. Experiments in continuous speech recognition have shown that these models decrease perplexity and improve recognition accuracy in comparison with conventional dialog-uniform stochastic language models.