A Novel Robust MFCC Extraction Method Using Sample-ISOMAP for Speech Recognition

According to the nonlinear characteristic of the speech signal, this paper presents a novel robust MFCC extraction method using sample-ISOMAP. ISOMAP is a nonlinear dimensionality reduction method based on the theory of manifold, it can reveal the meaningful low-dimensional structure hidden in the high-dimensional observations. In the proposed method, ISOMAP is first applied for calculating the non-linear mapping matrix which comes from the consistency mixed matrix. The consistency mixed matrix is composed of the logarithm of Mel filter bank energies derived from the sample data. Then the non-linear mapping matrix is used to replace the DCT procedure in the classic MFCC method. Experiments based on the recognition system established by HTK3.3 and Aurora2.0 speech database show that the robustness of the proposed method is superior to the PCA-MFCC and MFCC methods, and the recognition rate has been notably raised under low SNRs.