Speech recognition using semi-hidden Markov models of multiple features
暂无分享，去创建一个
Semi-hidden Markov models (SHMMs) have been suggested and applied to isolated speaker-dependent E-set recognition. The SHMM differs from the conventionaJ hidden Markov model (HMM) in that its states can be classified into types. A function which detects signals corresponding to state types is thus included in the SHMMs and utilized to supervise the estimation of their parameters. This general structure is implemented in the recognition experiment as models with their states classified into stationary and transient types. The average recognition error rate is about 18.9% which compares favourably with the average of about 36.4% reported when using a dynamic time warping (DTW) recognition system by Lienard and Soong (ref 3) on an equivalent vocabulary. Tests using corresponding HMMs show similar results to that of the DTW system.
[1] L. R. Rabiner,et al. An introduction to the application of the theory of probabilistic functions of a Markov process to automatic speech recognition , 1983, The Bell System Technical Journal.
[2] L. R. Rabiner,et al. Recognition of isolated digits using hidden Markov models with continuous mixture densities , 1985, AT&T Technical Journal.
[3] B.-H. Juang,et al. Maximum-likelihood estimation for mixture multivariate stochastic observations of Markov chains , 1985, AT&T Technical Journal.