Sixth European Conference on Speech Communication and Technology
MFCC is widely used together with its delta and delta-delta features in the field of speech recognition based on HMM. MFCC is designed to apply DCT to the MF output. We propose in this paper to employ KL transformation instead of DCT, because it can reflect the statistics of speech data more precisely. MFCCis the compressed feature of the log MFso that some detailed features seem to be lost. In this sense, we propose to compute the delta and delta-delta feature on the MF, and apply the KL transformation to a set of MF, its delta and delta-delta features.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Tokuhira, M. / Ariki, Y. (1999): "Effectiveness of KL-transformation in spectral delta expansion", In EUROSPEECH'99, 359-362.