An approach for training subspace distribution clustering HMM

In this paper, a new approach to train subspace distribution clustering HMM (SDCHMM) is described. With the multi-correlation coefficient and the Bhattacharyya distance, this approach is used for dividing the acoustical observation vector space, clustering subspace Gaussian distribution and getting subspace Gaussian prototypes. To evaluate the performance of the SDCHMM recognizer, a series of speaker-independent experiments are run to recognize Chinese digits. In comparison to a continuous density HMM (CDHMM) recognizer, a SDCHMM recognizer achieves 2- to 10-fold reduction in parameters requirement for acoustic models, and runs 20% - 25% faster without any loss of recognition accuracy.