Due to the short length, diversity, openness and colloquialism characteristic of out-of-domain (OOD) utterances, dialogue act (DA) recognition for OOD utterances in restricted domain spoken dialogue system remains a great challenge. This paper tackles this problem by proposing an effective DA recognition method using hybrid convolutional neural network (CNN) and random forest (RF). CNN acts as a feature extractor using Chinese character embeddings as original features. Then, RF classifier is trained on the learned features to perform DA recognition. The training dataset comes from a corpus of manually annotated Chinese OOD utterances, extracted from approximately 1500 conversations in a dialogue system which serves as a mobile phone recommendation assistant. We performed an evaluation on DA recognition for OOD utterances by logging the interactions of 15 subjects with the system. Experiment results justify the well performance of the proposed method.