Use of Kohonen self-organising feature maps for HMM parameter smoothing in speech recognition

The authors present a new method for smoothing the parameters of hidden Markov models (HMMs) which produces improved speech recognition results when only a limited amount of training data is available. The method uses the Kohonen self-organising feature map (KSOFM) as a clustering technique in codebook design for discrete HMMs. Neighbouring information provided by two-dimensional or three-dimensional KSOFM is used to smooth HMM symbol emission probabilities. Tested on a multispeaker isolated digit recognition task, the proposed smoothing method is shown to be robust. Speech recognition results are compared with those obtained by floor smoothing. Parzen smoothing and fuzzy vector quantisation. The major advantage of using a KSOFM as the vector quantiser in a HMM-based speech recogniser is that computational cost in both training and recognition stages can be reduced by using two speedup methods proposed in the paper