Estimation of posterior probabilities of discrete speech units - phonemes - has significant importance in basic speech processing research. The estimates are used in feature extraction (posterior features), phonotactic models for language recognition, generation of phoneme lattices for keyword spotting, and in other applications. The goal of this project is to create a fast and reliable system for estimation of such posterior probabilities that would allow to decrease error rates of the target systems. The project will deal with feature extraction, discriminative transforms, architectures of classifiers and techniques of training. The quality will be assessed mainly in international evaluations organized by US National Institute of Standards and Technology (NIST). (en)
tvorba rychlého a spolehlivého systému pro odhad posteriorních pravděpodobností fonémů