Inference of variable-length linguistic and acoustic units by multigrams

The efficiency of pattern recognition algorithms is highly conditioned to a proper definition of the patterns assumed to structure the data. The multigram model provides a statistical tool to retrieve sequential variable-length regularities within streams of data. In this paper, we present a general formulation of the model, applicable to single or multiple parallel strings of data having either discrete or continuous values. The model is first assessed to derive an inventory of variable-length sequences of letters from text data, where all spaces between the words have been removed. It turns out that the sequences of letters inferred during this fully unsupervised procedure clearly relate to the morphological structure of the text. The model is then used to infer a set of variable-length acoustic units, directly from speech data. Speech files containing examples of acoustic units are provided along with this paper in order to illustrate their consistency from an auditory point of view. We also report experiments using these acoustically defined units for continuous speech recognition.