Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

An Enhanced Interpolation Technique for Context-Specific Probability Estimation in Speech and Language Modelling

Fergus R. McInnes

Centre for Speech Technology Research, University of Edinburgh, Edinburgh, Scotland, UK

A recurring problem in speech and language modelling is the estimation of context-specific probability distributions from sparse data. Robust estimates can be obtained by interpolating the probability estimates obtained from context-specific statistics with more general ones. An interpolation technique is described here which, is based on a least-squares weighting formula but with deleted estimation incorporated to optimise its parameters. Perplexity results are given for various statistical language models incorporating this interpolation technique.

Full Paper

Bibliographic reference.  McInnes, Fergus R. (1992): "An enhanced interpolation technique for context-specific probability estimation in speech and language modelling", In ICSLP-1992, 1491-1494.