Since long, the use of contextual features has been shown to improve the recognition scores: use of numerical estimations of speed and acceleration appended to the current feature vectors, predictive HMM or neural networks. All these implementations are particular case of FIR filtering of feature trajectories. This paper presents a new approach where the characteristics of filters are trained together with the HMM parameters resulting in improvements of the recognition in first tests. Reestimation formulas for the cut-off frequencies of ideal LP-filters are derived as well for the impulse response coefficients of a general FIR LP-filter. Filters can be either common to all feature vectors or dedicated to a given entry or a given HMM state.
Cite as: Wellekens, C.J. (1998) Enhanced ASR by acoustic feature filtering. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0272, doi: 10.21437/ICSLP.1998-194
@inproceedings{wellekens98b_icslp,
author={Christian J. Wellekens},
title={{Enhanced ASR by acoustic feature filtering}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0272},
doi={10.21437/ICSLP.1998-194}
}