Speaker-independent isolated word recognition using a one-pass analysis

The system described in this paper is an outgrowth of one described in [1] and is a combination of deterministic and probabilistic speech recognition strategies. It is intended for telephone quality speech and possible single chip implementation with memory requirements of 256 bytes for processing (RAM) and 8k bytes for template storage and programs (ROM). Two major modifications for effecting reduction in memory requirements are described. The first is the use of a one-pass analysis algorithm for processing an utterance. The second is an approach for generation of uniform segment prototypes. Modification of the original system for telephone quality speech included obvious changes in filter bank analysis and parameterization of each phonetically classified segment with more emphasis on the voiced portions of utterances. The system was evaluated on a ten digit vocabulary (21 talkers, over 1000 tokens) and showed an accuracy of 94.4% with a 3.1% error rate and 2.5% rejection rate.