This paper presents a novel approach for speech emotion recognition named multilevel locality preserving K-SVD (MLP-KSVD). The proposed MLP-KSVD is designed to use with the coupled features, which are low-level and high-level features. In this study, the spectrogram and the Mel-spectrogram are regarded as the low-level and the high-level features, respectively. Our Idea is that the acoustic information behind the low-level feature contains more implicit information than those of the high-level feature. In contrast, the high-level feature even has lower dimensions but has the relatively explicit information. To explore collaboratively the advantages of the coupled features, the concept of coupled dictionary learning is applied to construct the dictionary. To preserve the geometric structure of low-level feature signals, a locality term is incorporated into our objective function. The experimental results on the extended MHMC dataset show that the proposed method outperforms all baselines including SRC, CRC, D-KSVD, LC-KSVD, and LP-KSVD.