Different methods for reducing parameters in a Gaussian mixture model (GMM) for text-independent speaker verification are investigated in this paper. The number of parameters is directly related to the memory requirement. Reducing the parameters is important in environments with limited memory resources or limited bandwidth for data transmission.
In contrast to standard approaches such as reducing the number of mixture components in the GMM or the dimension of the acoustic space, speaker specific parameters are selected from a global parameter set. Experiments reveal a small performance degradation when only a few parameters are chosen. Reducing the number of parameters to 25% of the original count gives a slightly better performance compared to a four times smaller global parameter set.
Cite as: Auckenthaler, R., Carey, M., Maso, J. (2000) Parameter reduction in a text-independent speaker verification system. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 3, 989-992, doi: 10.21437/ICSLP.2000-701
@inproceedings{auckenthaler00_icslp,
author={Roland Auckenthaler and Michael Carey and John Maso},
title={{Parameter reduction in a text-independent speaker verification system}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 3, 989-992},
doi={10.21437/ICSLP.2000-701}
}