Evaluation of the acoustic decoder of the `KEAL´ speech recognition system

KEAL, a continuous speech recognition system, is discussed. A C language version has been written in order to clarify the software architecture, to make possible the continuous upgrading and to facilitate a systematic and objective assessment. This has allowed the realization of larger scale experiments, and the measurements of KEAL´s qualities and drawbacks on various applications, with a reasonable number of different speakers. KEAL is one of the few examples of a complete system that has been run on a significant amount of real data. The authors review the organization of KEAL´s modules, give the results of its evaluation at the acoustic-phonetic level, and draw conclusions about its future evaluations