Abstract:
Word-internal context-dependent phoneme models, such as triphones, are used to create Hidden Markov Models (HMMs) for speech recognition. The large number of triphones results the excessive number of model parameters to be trained the HMMs. In order to reduce the number of model parameters, created triphones can be tied together to enhance the quality and robustness of HMMs. Data driven clustering or decision tree state clustering can be used to tie triphones to share the same set of parameters.