Speech recognition using localized affine invariant features

This paper proposes localized affine invariant features (LAIFs) for speaker-independent automatic speech recognition. The LAIFs can be calculated directly from data sequences. As speaker variations can be approximated well by affine transform in a cepstral space, the LAIFs can provide robust features with respect to those variations. This fact inspires us to expect that the use of the LAIFs should improve the recognition performance especially when no training data is available for speaker normalization or adaptation. To verify this expectation, we apply LAIFs for isolated word recognition. The experimental results show that the combination of LAIFs with MFCC or MFCC+∆MFCC can lead to higher performances than MFCC or MFCC+∆MFCC only. Especially in mismatched conditions, MFCC+∆MFCC+LAIFs can reduce the error rates by 37% when compared to MFCC+∆MFCC only.