Log-add approximation on-line noise compensation method adopting robust decision rules for robust digits recognition
Lists
Only you see this
Abstract
In this paper, we present a model-based noise compensation method. The method is a cascade of the Log-Add approximation noise compensation and robust decision rules. Log-Add approximation is extended to MFCC2, a MFCC (Mel-scaled frequency cepstral coefficient), like speech feature, which uses linear frequency amplitude and auditory-based preemphasis. A first pass of the method is used to obtain compensated models. In the second pass, a robust decision rule is adopted to eliminate mean prediction bias to some extent. Validity of the on-line noise compensation method is tested on TI digits recognition in several noise. Experiments show that the noise compensation method can effectively improve system performance in noisy environments. Also, we show that a proper combination of the robust decision rules with the Log-Add noise compensation can further improve system performance in noise.
BibTeX
@inproceedings{yao2000log,
author = {Kaisheng Yao and Bertram E. Shi and Pascale Fung and Zhigang Cao},
title = {{Log-add approximation on-line noise compensation method adopting robust decision rules for robust digits recognition}},
booktitle = {Rare},
year = {2000},
}Loading related content...