9104984 Cole This is the third year funding of a three-year continuing award. The ability to perform fine phonetic distinctions is a fundamental unsolved problem in computer speech recognition. Spoken letter recognition has practical applications and requires discrimination of acoustically similar words. An existing system will be improved in the following areas: (a) isolated letter recognition, (b) multiple letter recognition (sequences separated by brief pauses), and (c) continuous letter recognition (naturally spelled words). The result of this research will be a highly accurate, speaker-independent system that retrieves words from spellings. Strategies for improving isolated letter recognition include improving the signal representation and the broad category segmentation algorithm. Better letter segmentation strategies will be developed for those cases when the speaker does not pause as requested. Two-level classification will allow consideration of all possible letter sequences in conjunction with discrimination of whole-letter classification as in isolated letter recognition systems. Coarticulation models will be introduced to recognize between phoneme transitions, including dynamically adjusted duration models and boundary detection.
