Decoding optimal state sequence with smooth state likelihoods

A novel algorithm that allows the decoding of hidden Markov model (HMM) state sequences while constraining the state likelihoods to be more uniform is presented. In HMM-based speech recognizers, the decoded optimal state sequence is restricted by the HMM topology and the grammar. Thus, the most likely state sequence derived by the Viterbi algorithm can be influenced by a few states with very high likelihoods-often resulting in recognition errors. This paper presents a method for decoding state sequences with less volatile state probabilities by introducing penalties proportional to the difference of the current state likelihood and the highest state likelihood for the particular time frame. These penalties are added to the cumulative likelihoods in the Viterbi forward path at every time frame. This technique, referred to as the smooth state likelihood decoding algorithm (SSLDA), reduced recognition error-rates substantially on connected digit tests performed on two speech databases derived from field trials. The error rate was reduced by more than 40% on the one database and more than 60% on the other field trial database for variable length digit strings