A variable duration acoustic segment HMM for hard-to-recognize words and phrases

The authors consider how hidden Markov models (HMM) can be modified to accommodate a segment-based representation, and how word and subword models can be combined to improve recognition performance. The investigation is conducted within the context of a system that attempts to spot confusable words and phrases in the VOYAGER continuous speech corpus. Specifically, the authors describe how segments of varying duration are determined such that measurements for estimating the model parameters can be made on these segments. The word-spotting system is also described in detail.<<ETX>>