In this paper, we propose to use the missing data theory to allow the reconstruction of missing spectro-temporal parameters in the framework of hybrid HMM/ANN systems. A simple signal-to-noise ratio estimator is used to automatically detect the components that are unavailable or corrupted by noise (missing components). A limited number of multidimensional gaussian distributions are then used to reconstruct those missing components solely on the basis of the present data. The reconstructed vectors are then used as input to an artificial neural network estimating the HMM state probabilities. Continuous speech recognition experiments have been done on filtered speech. In this case, filtered components carry few or no information at all, and hence, should probably be ignored. The results presented in this paper illustrate this point of view. Complementary experiments also suggest the interest of the proposed approach in the case of noisy speech.
Cite as: Dupont, S. (1998) Missing data reconstruction for robust automatic speech recognition in the framework of hybrid HMM/ANN systems. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0581, doi: 10.21437/ICSLP.1998-323
@inproceedings{dupont98b_icslp,
author={Stéphane Dupont},
title={{Missing data reconstruction for robust automatic speech recognition in the framework of hybrid HMM/ANN systems}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0581},
doi={10.21437/ICSLP.1998-323}
}