Experimental Evaluation of Algorithms for Connected Speech Recognition Using Hidden Markov Models.

Current Automatic Speech Recognition devices attempt to solve the connected word recognition problem by assuming that an unknown phrase is the output of a sequence of statistical word-models. Typically, these models are constructed using examples of words spoken in isolation; however, the acoustic patterns corresponding to words as they occur in fluent speech are quite different from those representing the same words spoken in isolation, and so the use in speech recognisers of models based on isolated utterances severely limits the performance of such devices. A method of extracting training utterances from fluent speech and constructing Hidden Markov Models (HMMs) from these templates, known as Embedded Training, is investigated here, in conjunction with a two-level algorithm for connected word recognition. The effects on recognition performance of various HMM training procedures are discussed, and experimental results are presented