A novel two-level model of timing in speech is described. In this model, phoneme durations are first calculated at the lower-level of segmental characteristics and contextual effects, and at a second stage, the already predicted durations are modified to reflect the higher-level of rhythmic and structural organisation of the utterance. The model is based on results obtained from analysis of isolated words for the first level and results from continuous speech analysis for the second one. The complete model can be used in both isolated word and continuous speech recognition systems, and Text-To-Speech systems. The model has been implemented in the Greek language.

Keywords: Duration models, Text-To-Speech systems, Isolated Word Recognition systems, Greek language