Segment-Based Acoustic Models for Continuous Speech Recognition

ity or acoustic observations conditioned on the state in Tied-mixture (or semi-continuous) distributions are an imhidden-Markov models (11MM), or for the case of the portant tool for acoustic modeling, used in many highSSM, conditioned on a region of the model. Some of the performance speech recognition systems today. This paper options that have been investigated include discrete dispiovides a survey of the work in this area, outlining the tributions based on vector quantization, as well as Gausdifferent options available for tied mixture modeling, introsian, Gaussian mixture and tied-Gaussian mixture disducing algorithms for reducing training time, and providtributions. In tied-mixture modeling, distributions are ing experimental results assessing the trade-offs for speakermodeled as a mixture of continuous densities, but unlike independent recognition on the Resource Management task. ordinary, non-tied mixtures, rather than estimating the Additionally, we describe an extension of tied mixtures to component Gaussian densities separately, each mixture segment-level distributions, is constrained to share the same component densities with only the weights differing. The probability density