Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Smoothing Hidden Markov Models ay Means of a Self Organizing Feature Map

E. Monte, Josť B. Marino, Eduardo LLeida

E.T.S.E. Telecomunicacio, Barcelona, Spain

This paper proposes a method for smoothing the Hidden Markov Models (HMM) with the VQ done by means of the Self Organising Feature Maps (SOFM). The use SOFM gives rise to a special property of the probability of emission matrix of the HMM. This property is that when ordering the probability of emission matrix following the order of the SOFM; neighbouring symbols will have similar probabilities. In order to smooth the HMM we propose to filter the probability of emission matrix by a filter that makes use of this property. We also compare this method with another method for smoothing the HMM; the coocurrence method. The recognition rate improvement achieved by the method that we propose is better than the recognition rate obtained by means of the coocurrence method.

Full Paper

Bibliographic reference.  Monte, E. / Marino, Josť B. / LLeida, Eduardo (1992): "Smoothing hidden Markov models ay means of a self organizing feature map", In ICSLP-1992, 535-538.