Transient part recognition for continuous speech using transition spotting

Describes a recognition system for the transient parts of continuous speech. The algorithm uses DTW as distortion measure, and applies it starting at each frame of the unknown signal. The localization of the transient parts in the speech signal is made using a spotting transient part dictionary, as well as with a full diphone one. The output of this system is a lattice of hypothesis among which 65% of the transitions are correctly spotted and recognized. A complete system, using a steady part recognizer has also been designed.<
>