The invention provides a speed emotion identification method based on a depth belief network and a kernel non-linear PSVM, and relates to the emotion identification technical field; the method mainlycomprises the following steps: emotion signal obtaining; emotion signal preprocessing; emotion characteristic parameter extraction; emotion characteristic parameter integration; emotion identification. The method firstly uses the depth belief network to automatically extract new speech emotion characteristics from preprocessed emotion signals, integrates the new characteristics with conventional speech emotion characteristics so as to obtain an integration characteristic vector, uses the integration characteristic vector as the input of the kernel non-linear PSVM, finally uses a majority voting principle to vote for output results of all kernel non-linear PSVM classifiers, thus obtaining the final identification result, and identifying four basic human emotions such as angry, happy, sad and marvel. The method can fully express the advantages of a depth learning algorithm and the PSVM classifiers, thus enabling the whole emotion information identification process to more approach to human emotion identification, and improving the emotion identification accuracy.