Maximum likelihood multiple projection schemes for hidden Markov models
Mark Gales · Cambridge University Engineering Department Publications Database · 1999
In many complex pattern recognition tasks the choice of a \\good" feature space in which to model the data varies depending on the signal content. Multiple feature subspaces can be used, for example phone dependent subspaces in speech recognition. Handling multiple subspaces whilst still maintaining meaningful likelihood comparisons between classes is a complex problem. This paper views this multiple projection problem in a maximum likelihood framework. Two forms of multiple projection schemes are described in terms of tying a standard semi-tied covariance matrix system. The rst results in model complexity control over the dimensions, whilst selecting the appropriate feature space for that complexity. The second may be viewed as a simultaneous projection and optimisation of a set of semi-tied transforms. These two schemes are compared on a large vocabulary task. Varying the complexity over the dimension was found to slightly increase speed and reduce model size without degrading the pe...