In the last few years, Support Vector Machine classifiers have been shown to give results comparable, or better, than Hidden Markov Models for a variety of tasks involving variable length sequential data. This type of data arises naturally in the fields of bioinformatics, text categorization and automatic speech recognition. In particular, in a previous work it was shown that certain string kernels gave a classification performance comparable to discrete Hidden Markov Models on an isolated Spanish digit recognition task. It is known that speech recognition degrades, often quite severely, when noise is present, and it is interesting to ask whether Support Vector Machines with string kernels continue to give a similar proficiency to discrete Hidden Markov Models in this context. In the present paper, this question is explored by considering the performance of Support Vector Machines with string kernels on the same isolated Spanish digit recognition task in which the speech data has been corrupted with different types of noise. Specifically, white noise and speech babble from the NOISEX-92 database. Results of these experiments are given.
Cite as: Goddard, J., Martinez, A.E., Martinez, F.M., Rufiner, H.L. (2004) Noisy speech recognition using string kernels. Proc. 9th Conference on Speech and Computer (SPECOM 2004), 89-94
@inproceedings{goddard04_specom,
author={J. Goddard and A. E. Martinez and F. M. Martinez and H. L. Rufiner},
title={{Noisy speech recognition using string kernels}},
year=2004,
booktitle={Proc. 9th Conference on Speech and Computer (SPECOM 2004)},
pages={89--94}
}