The use of a distribution-clustering technique in HMM-based continuous-speech recognition

Hidden Markov modelling is one of the most powerful and popular representations of acoustic phenomena in isolated-word or continuous speech recognition. In this case, an essential challenge is to achieve a trade-off between the complexity of the acoustic models and their trainability. In order to do so, the authors have defined a shared-distribution approach in their HMM-based continuous-speech recognizer. In this clustering algorithm the distortion measure between two distributions is only based on the weights of Gaussian mixtures rather than on all parameters of the distributions. Experimental results on the ATIS task show that their shared-distribution approach increased by 6% the word accuracy rate in comparison with the baseline system