MicroRNAs(miRNAs) are short non-coding RNAs that play important regulatory roles in both animals and plants. While the first miRNAs were discovered using experimental methods, experimental miRNA identification remains technically challenging and incomplete. Hence, computational approaches are a natural choice to complement experimental approaches to miRNA gene identification. A de novo miRNA precursor prediction method was proposed. In constructing the recognition model, both primary sequence and secondary structure were combined into an input sequence through encoding, and the input space was mapped into a feature space via k-gram method. After applying feature selection, those selected features was used to construct SVM-based models for the recognition of miRNA precursors. In the mean time, the method was compared with the HMM learning method. Experimental results show that the method outperforms HMM. The reason is that microRNAs are so short that it is not easy for HMM model to capture the signals for differentiating the genuine microRNAs from those pseudo-microRNA genes. From features selected, it was found that they are mostly come from the primary and secondary structure of microRNAs. This phenomenon may tell us to put more efforts in the microRNAs themselves in designing computational method before we fully understand the transcription mechanism of microRNA biologically.