Probabilistic-trajectory segmental HMMs

摘要: Abstract “Segmental hidden Markov models” (SHMMs) are intended to overcome important speech-modelling limitations of the conventional-HMM approach by representing sequences (or segments) features and incorporating concept trajectories describe how change over time. A novel feature presented in this paper is thatextra-segmentalvariability between different examples a sub-phonemic speech segment modelled separately fromintra-segmentalvariability within any one example. The extra-segmental component model represented terms variability trajectory parameters, these models therefore referred as “probabilistic-trajectory segmental HMMs” (PTSHMMs). This presents theory PTSHMMs using linear description characterized slope mid-point theoretical experimental comparisons types PTSHMMs, simpler SHMMs conventional HMMs. Experiments have demonstrated that, for given set, PTSHMM can substantially reduce error rate comparison with HMM, both connected-digit recognition task phonetic classification task. Performance benefits been from additionally modelling parameter.