Tech. Rep. - Ref. No. 8904-14 - February 1989
The prototype of an automatic connected speech recognition system is presented, in which a simple but robust segmentation procedure is used to find acoustic segments, not necessarily corresponding to words. These segments are terminal symbols of a regular grammar used during the recognition phase to drive the search. This acoustic grammar is generated, in semi-automatic way, starting from an equivalent grammar whose terminal symbols are words. With respect to the search space complexity, the resulting recognition task is comparable with an isolated word recognition task. Details of the whole system, the software environment, the algorithm for the automatic grammar generation, and the performance referred to a rather simple application are described and discussed.