14thAnnual Conference of the International Speech Communication Association

Lyon, France
August 25-29, 2013

Augmented Conditional Random Fields Modeling Based on Discriminatively Trained Features

Yasser Hifny

University of Helwan, Egypt

Augmented Conditional Random Fields (ACRFs) are undirected graphical models that maintain the Markov properties of Hidden Markov Models (HMMs), formulated using the maximum entropy (MaxEnt) principle. ACRFs incorporate acoustic context information into an augmented space in order to model the sequential phenomena of the speech signal. The augmented space is constructed using Gaussian activation functions representing the dense regions in the observation space. These activation functions are estimated using the Expectation-Maximization (EM) algorithm. Alternatively, the activation functions can be estimated using a discriminative objective function. Hence, the ACRFs are fed with discriminative features. In this paper, we show that ACRFs recognition results improve if the activation functions are estimated using the Minimum Phone Error (MPE) discriminative criterion.

Full Paper

Bibliographic reference.  Hifny, Yasser (2013): "Augmented conditional random fields modeling based on discriminatively trained features", In INTERSPEECH-2013, 2341-2344.