Increasing the mixture components of non-uniform HMM structures based on a variational Bayesian approach.

摘要: We propose using the Variational Bayesian (VB) approach for automatically creating non-uniform, context-dependent HMM topologies. Although Maximum Likelihood (ML) criterion is generally used to create topologies, it has an overfitting problem. Recently, avoid this problem, VB been applied acoustic models speech recognition. introduce Successive State Splitting (SSS) algorithm, which can both contextual and temporal variations HMMs. Experimental results show that proposed method a more efficient model than original method. Furthermore, we evaluated increase number of mixture components by considering structures. The obtained best performance with smaller in comparison ML based methods.