A pattern classification technique for speech recognition

A description is given of an unusual pattern recognition technique which has been used in an experimental speech recognition system. Preliminary results obtained using this technique are reported. The speech analyzer produces a multichannel ternary signal at its output, which is the short term digital autocorrelation function of the input signal. This output is sampled at regular intervals and this sampled information is transferred to a computer. A new pattern recognition technique is proposed that avoids the exhaustive comparison process associated with pattern matching. The technique is similar to a tree-structured process in that decisions are taken that exclude certain master patterns from further processing as it becomes apparent that these are sufficiently dissimilar to the unknown pattern. However, retracing within the structure and the substitution of an alternative path are permitted if the current path appears unlikely to lead to a correct decision. Some preliminary results obtained using this technique are described. These show that a performance very similar to that obtained from the exhaustive comparison process is attainable with a significant saving in computational effort. The effect of varying certain parameters within the recognition process is also considered and some preliminary optimization of parameter values is reported.