Signalrepresentasjoner for automatisk talegjenkjenning

Abstract

In this report we give an overwiev of methods for front-end processing of speech signals for automatic speech recognition (ASR) that are described in the litterature. The most common representation of speech in this context seems to be mel-frequency cepstral coeficient (MFCC) with delta- and double-delta coefficients, usually combined with cepstral mean normalization (CMN). Other representations include perceptual linear prediction (PLP) and linear prediction cepstral coefficients (LPCC).