A new method, named linear prediction with linear extrapolation has been proposed in the past, which aims at modifying conventional linear prediction especially for speech coding applications. The basic idea is to reformulate the computation of linear prediction so that an optimal FIR-predictor of order 2p could be determined from p numerical values. In this work, we extend the above method to generate the cepstral features using the compressed LPC parameters and use the new feature for accurate speech recognition. Peliminary results on Tamil connected digit recognition task demonstrated that the cepstral features derived from the new approach yield more accurate modeling of speech spectra and provides better discrimination among different speech classes.
Cite as: Chengalvarayan, R. (2000) Use of linear extrapolation based linear predictive cepstral features (LE-LPCC) for Tamil speech recognition. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 2, 903-906, doi: 10.21437/ICSLP.2000-416
@inproceedings{chengalvarayan00b_icslp,
author={Rathinavelu Chengalvarayan},
title={{Use of linear extrapolation based linear predictive cepstral features (LE-LPCC) for Tamil speech recognition}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 2, 903-906},
doi={10.21437/ICSLP.2000-416}
}