MODELLING FOR CONT INUOUS-SPEECH RECOGNITION

Clustering techniques have been integ rated at different levels into the training procedure of a continuous-densi ty hidden Markov model (HMM ) speech recognizer. These clusterin g techniques can be used in two ways. First acous­ tically similar states are tied together. It will help to reduce the number of parameters but also allow to train otherwise rarely seen states together with more robust ones (state­ tying ). Secondly densities are clustered across states, this reduces the number of densities while at the same time keeping the best performa nces of our recognizer (density­ clustering). We have applied these techniques both to word­ based small-vocabulary and phoneme-based large-vocabula­ ry recognition tasks. On the WSJ task, we could achieve a reduction of the word error rate by 7%. On the TljNIST­ connected digit task, the number of parameters was reduced by a. fa.ctor 2-3 while keeping the same string error rate.