In designing classifiers for automatic speech recognitions, one of the problems the user faces is to cope with an unwanted variability in the environment such as changes in the speaker or the acoustics. To overcome this problem, various adaptation schemes have been proposed in the literature. In this short paper, rather than selecting a single acoustic model as being representative of a category, we adaptively find the optimal or near-optimal number of hidden Markov models during the Baum-Welch (BW) learning process through splitting and merging operations. This scheme is based on incorporating the split-merge operations into the HMM parameter re-estimation process of the BW algorithm. In the splitting phase, an acoustic model is divided into two sub-models based on a suitable criterion. On the other hand, in the merging phase, two models are combined into a single one. The experimental results demonstrate that the proposed mechanism can efficiently resolve the problem by adjusting the number of acoustic models while increasing the classification accuracy. The results also demonstrate that the advantage gained in the case of multi-modally distributed data sets is significant.
展开▼
