Protein fold recognition has been the focus of computational biologists for many years. In order to map a protein primary structure to its correct 3D fold, we introduce in this paper a machine learning paradigm that we entitled "structural hidden Markov model" (SHMM). We show how the concept of SHMM can efficiently use the protein secondary structure during the fold recognition task. Experimental results showed that the SHMMoutperforms the SVM with a 6% improvement in the average accuracy. However, because in this application the two classifiers are not correlated, therefore their combination based on the highest rank criterion boosted the SHMM average accuracy with 10%.