An adaptive memory technique for speech recognition

A new method for registering pattern features in electronic templates has been devised which does not use error-correction routines. It is easy to train, does not require additional training when categories are added, and uses low-cost, reliable, binarystorage devices. Three experimental models have been developed and tested. The first, using fixed assignments of electronic templates to categories, is suitable for adapting to an arbitrary library of 16 words for one speaker at a time. The second, using an adaptive assignment of electronic templates to categories, is adaptable to several speakers simultaneously or to different languages. The third is a new system for connected-word recognition, using adaptive assignment of templates according to size.