[Univ of Cambridge]alt[Dept of Engineering]

MIL Speech Seminars 2003-2004

The MIL Speech Seminar series schedule for Michaelmas Term 2002 was as follows:

November 19th 2002 Marcus Tomalin (MIL) The Art of the SVR Speech Seminar One of the purposes of the SVR Speech seminars is to enable SVR Speech researchers to present their work to other members of the group. Since there are so many new people joining the group this year (esp. RAs and PhDs), all of whom will be expected to give a seminar presentation at some point during their time at CUED, it is appropriate to offer an introductory presentation designed to provide information and advice concerning seminar preparation. The intention is to ensure that, in the months ahead, ignorance of practicalities and procedure cannot be cited as a valid reason for not offering a seminar presentation. You have been warned :)
November 25th 2002 Antti-Veikko Rosti (MIL) Linear Gaussian Models for Speech Recognition The aim of this talk is to introduce a group of state-space models, called linear Gaussian models (LGM), in a general framework. In comparison to standard HMM, which also belongs to LGMs, some forms can improve known shortcomings in modelling spatial and temporal correlation. Models based on factor analysis observation process and both, HMM and linear first-order Gauss-Markov state evolution processes, are introduced with applications in speech recognition. The first, factor analysed HMM (FAHMM) generalises into many standard systems such as semi-tied full covariance HMM and shared factor analysis. The second model, switching linear dynamical system is an extension to a stochastic segment model to include inter-segment information.
December 10th 2002 Dan Povey (MIL) Discriminative Training for Speech Recognition In last few years, the use of discriminative training techniques in the context of large vocabulary speech recognition systems has become an area of active research, and this seminar will discuss various aspects of the use of discriminative training in such systems. The Maximum Mutual Information (MMI) and Minimum Phone Error (MPE) objective functions will be presented and discussed with reference to recent experimental results. Also the use of lattices and the Extended Baum-Welch algorithm for optimisation will be described, as will the weak-sense and strong-sense auxiliary functions. Finally, the benefits of discriminative training across various speech corpora will be assessed. This seminar should be of particular interest to anyone who may be using discriminative techniques in the near future, but who lacks a clear and comprehensive understanding of the underlying theoretical and implementational issues that are attendant upon the use of such techniques.