Anchor point detection for continuous speech recognition in Spanish: the spotting of phonetic events

Several techniques have been used to constrain the search space in an HMM based continuous speech recognition system (grammar, beam search, etc.) in order to reduce computation without significant lose in performance (H. Ney, 1993). The use of anchor points is a state-of-the-art technique that has already been used in some systems (D. Lubensky, 1987; H. Murveit et al., 1993; M. Adda-Decker, 1989). This approach has lead us to a bottom-up strategy in a continuous speech recognition system in which the first module performs the spotting of predefined phonetic events and the second module uses them as anchor points in order to guide the HMM based recognition task. The paper describes the definition of phonetic events for Spanish (based on expert knowledge of language) and the algorithms used for their detection and classification. Performance figures are presented