Performance through consistency: connectionist large vocabulary continuous speech recognition

It is shown that the performance of a connectionist speech recognition system can be improved by resolving inconsistencies in its design. Specifically, by introducing word level training into a TDNN (time-delay neural network) phoneme classifier (thus defining a multistate time-delay neural network or MS-TDNN), the training and testing criteria become consistent, enhancing the system´s word recognition accuracy. The author applied this MS-TDNN architecture to the task of large vocabulary continuous speech recognition, and found that it outperforms all other systems that have been evaluated on the CMU (Carnegie Mellon University) Conference Registration database. In addition, preliminary results suggest that the MS-TDNN may perform well on the large vocabulary Resource Management database, using a relatively small number of free parameters.<>