GENONES - GENERALIZED MIXTURE TYING IN CONTINUOUS HIDDEN MARKOV MODEL-BASED SPEECH RECOGNIZERS

An algorithm is proposed that achieves a good tradeoff between modelin g resolution and robustness by using a new, general scheme for tying o f mixture components in continuous mixture-density hidden Markov model (HMM)-based speech recognizers. The sets of HMM states that share the same mixture components are determined automatically using agglomerat ive clustering techniques. Experimental results on ARPA's Wall Street Journal corpus show that this scheme reduces errors by 25% over typica l tied-mixture systems. New fast algorithms for computing Gaussian lik elihoods-the most time-consuming aspect of continuous-density HMM syst ems-are also presented. These new algorithms significantly reduce the number of Gaussian densities that are evaluated with little or no impa ct on speech recognition accuracy.