Detection of phoneme boundary by mixed order Markov process

This paper introduces a new segmentation algorithm of speech signal based on Markov process. By assuming speech spectrum and segmentation boundary as a 1st and 3rd order Markov process, respectively, we convert the segmentation problem to an optimization problem. The mean field theory allows a MAP solution which can be implemented with sigmoid-like neurons. The experimental results show that our algorithm is superior to the previous method based on conditional cross entropy, in accuracy as well as in computational speed.