Abstract
We propose the use of discriminativetrainingbymeansofthe generalized probabilistic descent #GPD# algorithm to estimatehidden Markov model #HMM# stream exponents foraudio-visual speech recognition. Synchronized audio and visualfeatures are used to respectively train audio-only andvisual-only single-stream HMMs of identical topology bymaximum likelihood. A two-stream HMM is then obtainedby combining the two single-stream HMMs and introducingexponents that weigh the log-likelihood of each ...

This publication has 4 references indexed in Scilit: