A novel feature extraction using multiple acoustic feature planes for HMM-based speech recognition

This paper describes an attempt to extract multiple peripheral features of a point x(ti,fj) on a timespectrum (TS) pattern by observing n×n neighborhoods of the point, and to incorporate these peripheral features (MPFPs: multiple peripheral feature planes) into the feature extractor of a speech recognition system together with MFCC parameters. Two types of peripheral feature extractor, MPFP-KL and MPFP-LR, are proposed. MPFP-KL adopts the orthogonal bases extracted directly from speech data by using KLT of 7×7 blocks on TS patterns. In MPFP-LR, the upper two primal bases are selected and simplified in the form of ∆t-operator and ∆foperator obtained by linear regression calculation. MPFP-KL and MPFP-LR show significant improvements in comparison with the standard MFCC feature extractor in experiments with the HMM-based ASR system.