At the present time, one of the most important problem in large vocabulary continuous speech recognition is to achieve an optimum trade-off between acoustic models complexity and their trainability. In order to do so, we have defined a shared-distribution approach in our HMM-based continuous speech recognizer. In this clustering algorithm the distortion measure between two distributions is only based on the weights of gaussian mixture rather than all parameters of the distributions. Experimental results on the ATIS task show that our shared-distribution approach increased by 6% the word accuracy rate in comparison with our baseline system.
Bibliographic reference. Farhat, Azarshid / O'Shaughnessy, Douglas (1995): "A shared-distribution approach in a hidden Markov model-based continuous speech recognition system", In EUROSPEECH-1995, 1503-1506.