Sixth International Conference on Spoken Language Processing (ICSLP 2000)

Beijing, China
October 16-20, 2000

Multi-Group Mixture Weight HMM

Ming Li, Tiecheng Yu

Speech Processing Lab, Institute of Acoustics, Chinese Academy of Sciences Beijing, China

This paper presents a new modeling method of the continuous density Hidden Markov Model. As we know, speech signal is characterized by a hidden state sequence and each state is described by the mixture of weighted Gaussian density functions. Usually if we want to describe speech signal more precisely, we need to use more Gaussian functions for each state. But it will increase the computation significantly. On the other hand, the weight of each Gaussian component is the statistical average of Gaussian component probabilities for the whole training data. So it just can depict the average characteristics of speech signal. For some speech signal these weights are not proper in fact. Therefore, we propose Multi-group Mixture Weight HMM to solve this problem. In this kind of HMM, each state has several groups of mixture weight for the Gaussian components and it only needs very little additional computation. In our experiments, it achieved 12% reduction for errors.

Full Paper

Bibliographic reference.  Li, Ming / Yu, Tiecheng (2000): "Multi-group mixture weight HMM", In ICSLP-2000, vol.1, 290-292.