MCE-based training of subspace distribution clustering HMM

Abstract

For resource-limited platforms, the subspace distribution clustering hidden Markov model (SDCHMM) is better than the continuous density hidden Markov model (CDHMM) for its smaller storage and lower computations while maintaining a decent recognition performance. But the normal SDCHMM obtaining method does not ensure optimality in classifier design. In order to obtain an optimal classifier, a new SDCHMM training algorithm that adjusts the parameters of SDCHMM according to the minimum classification error (MCE) criterion is proposed in this paper. Our experimental results on TiDigits and RM tasks show the MCE-based SDCHMM training algorithm provides 15-80% word error rate reduction (WERR) compared with the normal SDCHMM that is converted from CDHMM.