Stabilized weighted linear prediction (SWLP) is a recently developed method to compute stable all-pole models of speech by applying temporal weighting of the residual energy. In this study, SWLP is used for spectrum estimation in the first stage of the MFCC computation. The resulting acoustic feature representation is tested in a speech recognition front-end in simulated noisy conditions. When compared to other spectrum estimation methods as a part of the MFCC framework, the proposed spectrum estimation method clearly outperforms the FFT (periodogram), linear prediction and minimum variance distortionless response (MVDR) methods in terms of noise robustness.
Cite as: Pohjalainen, J., Magi, C., Alku, P. (2008) Enhancing noise robustness in automatic speech recognition using stabilized weighted linear prediction (SWLP). Proc. ISCA ITRW on Speech Analysis and Processing for Knowledge Discovery, paper 002
@inproceedings{pohjalainen08_spkd,
author={Jouni Pohjalainen and Carlo Magi and Paavo Alku},
title={{Enhancing noise robustness in automatic speech recognition using stabilized weighted linear prediction (SWLP)}},
year=2008,
booktitle={Proc. ISCA ITRW on Speech Analysis and Processing for Knowledge Discovery},
pages={paper 002}
}