Research continues in the area of advanced speech recognition algorithms, architectures, and experimentation. Algorithms currently being tested include an ear-model-based filterbank for speech recognition analysis, a dynamic programming solution for connected speech recognition, and a DILEX approach for continuous digit recognition. New algorithms are being investigated which apply expert system technology and/or hidden Markov models or Markov random fields. Finally, a training paradigm for large-vocabulary discrete utterances recognition (thousands of utterances) is being developed using some of these same ideas. The largest real hardware project, Armstrong, a distributed multiprocessor, is in its very early stages of construction and consists of a unique network of some 100 M68010-based processors. Another phase of the program relates to the development of a hardware/software tool designed to be a high-quality intelligent test instrument for speech recognition systems.
