AVSP 2003 - International Conference on Audio-Visual Speech Processing

September 4-7, 2003
St. Jorioz, France

Why the FLMP Should Not be Applied to McGurk data ...or how to Better Compare Models in the Bayesian Framework

Jean-Luc Schwartz

Institut de la Communication Parlée, CNRS UMR 5009, INPG-Université Stendhal INPG, Grenoble, France

We come back on two criticisms addressed to the "Fuzzy-Logical Model of Perception" (FLMP): its assumed ability to fit any data, and the choice of a "fitting all" procedure rather than a "predicting AV from A and V" approach. We confirm the FLMP ability to fit random data in the "McGurk region", that is in all conditions including conflicting stimuli, thanks to the "0/0 trick". This is associated to a high instability of the rmse value in the optimal region. Then we introduce a Bayesian approach replacing fit by global likelihood. In this sounder framework, rmse instabilities decrease likelihood, hence the 0/0 trick is problematic for the FLMP. We conclude by indicating when the FLMP can be used in sound conditions: apart from the "McGurk" region.

Full Paper

Bibliographic reference.  Schwartz, Jean-Luc (2003): "Why the FLMP should not be applied to McGurk data ...or how to better compare models in the Bayesian framework", In AVSP 2003, 77-82.