Forensic data from long-term formant analysis were used as input to the GMM-UBM approach, which is a way of deriving Likelihood Ratios. Tests were performed running 22 same-speaker comparisons and 462 different-speaker comparisons from a corpus of anonymized casework data involving telephone-intercepted speech. In a first series of tests, the number of Gaussian modules for GMM-modeling was increased from 1 to 32. In a second series of tests the duration of formant input in the compared files was reduced from 10 seconds to 5 and then to 2.5. All tests were performed both without and with the use of MAP adaptation. Results were evaluated in terms of overall performance characteristics EER and Cllr and in terms of score distributions visualized as Tippett plots. The main goal of the study was to compare the use and non-use of MAP and to look at the practical forensic implications of the difference. Results show that in terms of overall performance characteristics there is little difference between the selection and de-selection of MAP. Tippett plot patterns however reveal strong differences. Application of MAP allows for more symmetric same- and different-speaker distributions and shows more robustness against duration reductions, both of which are forensically important.
Cite as: Jessen, M. (2021) MAP Adaptation Characteristics in Forensic Long-Term Formant Analysis. Proc. Interspeech 2021, 411-415, doi: 10.21437/Interspeech.2021-1697
@inproceedings{jessen21_interspeech,
author={Michael Jessen},
title={{MAP Adaptation Characteristics in Forensic Long-Term Formant Analysis}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={411--415},
doi={10.21437/Interspeech.2021-1697}
}