Kim Kwang Tae, Seo Jeong Il, Hong Jae Keun

The Transactions of the Korea Information Processing Society (1994 ~ 2000), Vol. 5, No. 2, pp. 537-544, Feb. 1998
10.3745/KIPSTE.1998.5.2.537,   PDF Download:

Abstract

It is made the speaker adaptation model by adjusting both the mean and the variance of the Gaussian state observation densities of a CDHMM to use the MAPE method. However, we can't use the MAPE method in ARHMM because the components of LPC vector are used as the feature vector of ARHMM. Therefore, in this paper, we propose a speaker adaption method of ARHMM to adapt the speaker adaptation model having one branch in each state after it is divided the input utterance, which is spoken to an adapted speaker, into states by Viterbi algorithm and then make a typical vector using modified k-means algorithm. In addition we have experimented another method in which each state is represented by several branches. If the training data is insufficient, this method is not proper to train. So we vary the number of branch in proportion to the number of frame stayed in each state, and make to absorb the characteristics of speaker's pronunciation speed and duration by using the distribution of the state duration adapted to the speaker. When testing 15-word Korean domestic name isolated word model, using the proposed method, the recognition performance was found to reduce the error rate of speaker-independent systems more than 50%.

Statistics
Cite this article