A Markov model containing state-conditioned second-order non-stationarity: application to speech recognition

摘要: Abstract In this paper we report our development of a new class hidden Markov models (HMMs) with each state characterized by time series model which is non-stationary up to the second order. A close-form solution for parameter estimation obtained based on EM algorithm and matrix-calculus implementation technique. first set evaluation experiments, adopt residual square sum, over states frames within bounds, as quantitative measure goodness fit between speech data. It observed that inclusion state-conditioned second-order non-stationarity, implemented use time-varying regression coefficients, has substantially greater effects reducing data-fitting error than increase terms while maintaining coefficients term constant. set, isolated-word recognition it found mix first-order non-stationarities consistently produces higher accuracy conventional, stationary-state HMMs.