Selective multi-path acoustic model based on database likelihoods

An efficient multi-path acoustic model based on database likelihoods for spontaneous speech recognition is presented. Although a multi-path phone HMM that has several mod­ els for di仔erent target in parallel is considered effective to express multi-style or speed-variant nature of spontaneous speech， assuming various model to match at every time for all phones may cause mismatch of unintended model， and spoil the model constraints. We propose defining a multi­ path model that has several different state resolutions only for the distortive phones selectively. The phone set is se­ lected through an analysis of the likelihoods and duration times of phone segments in a spoken dialogue co叩us using automatic viterbi alignment. Experiments on three testsets showed that our multi-path model based on the phone selec­ tion can achieve better accuracy than a simple single-path model， whereas a full multi-pa出model without phone se­ lection causes much degradation of accuracy.