Graphical models and automatic speech recognition
Jeff Bilmes · The Journal of the Acoustical Society of America · 2002
Graphical models (GMs) are a flexible statistical abstraction that has been successfully used to describe problems in a variety of different domains. Commonly used for ASR, hidden Markov models are only one example of the large space of models constituting GMs. Therefore, GMs are useful to understand existing ASR approaches and also offer a promising path towards novel techniques. In this work, several such ways are described including (1) using both directed and undirected GMs to represent sparse Gaussian and conditional Gaussian distributions, (2) GMs for representing information fusion and classifier combination, (3) GMs for representing hidden articulatory information in a speech signal, (4) structural discriminability where the graph structure itself is discriminative, and the difficulties that arise when learning discriminative structure (5) switching graph structures, where the graph may change dynamically, and (6) language modeling. The graphical model toolkit (GMTK), a software system for general graphical-model based speech recognition and time series analysis, will also be described, including a number of GMTK’s features that are specifically geared to ASR.