Efficient Mllr

The need for close to real time speech recognition has recently driven interest in fast LVCSR systems. Due to the time constraint, such systems often discard, where possible, sub-processes of the entire recognition process which demand relatively large amounts of computation and yield relatively small accuracy gains. This report focusses on such speed-accuracy tradeoffs with regard to speaker adaptation. A variety of techniques are used to reduce the compute time of a baseline adaptation system which uses mean-only maximum likelihood linear regression (MLLR). Use of a mixture component-level Viterbi alignment to accumulate adaptation statistics and least squares linear regression transform estimation are compared to the baseline techniques in terms of speed and accuracy. In the case of unsupervised adaptation, exploitation of word boundary information generated in an initial recognition pass is shown to further reduce adaptation time with no negative impact upon system accuracy. Again, in the unsupervised case, an illustration of how confidence scores can be used to simultaneously reduce adaptation time and improve accuracy is presented.