Optimized state-tying for triphone-based HMMs under training data deficiency

This paper deals with an optimization of state-tying for triphone-based HMM in the case of training data deficiency. The main goal is to analyse the importance of stopping threshold for criterial function in tree-based clustering. The log-likelihood measure was used as the criterial function, when a varying threshold with different sizes of training set was evaluated. Tied-state triphone HMMs with multiple Gaussian mixtures were trained under various setups. Realized experiments showed that the more complex AMs with less mixtures added could achieve better results that less complex models with more mixtures. The same conclusion was proved for even significantly reduced amount of training data.