Automatic speech recognition for large vocabularies

An isolated word recognition system for large vocabularies (1000 to 5000 words) with 98% recognition performance was developed. It was implemented on an array processor for real time requirements. The speech signal is described by short time autocorrelation functions. Short response times as well as high recognition accuracies are achieved by means of a hierarchical classification scheme. A fast preselection stage yields a small number of suitable word candidates to be considered for further classification. To that end a linear segmentation or a segmentation based on acoustic or phonetic cues was performed. High selectivity is obtained by using fine temporal resolution and nonlinear time alignment in the final classification step. By taking into account phonetically identical fragments of words, a distinction between highly confusable words can be made. Speaker adaptation for new system users is performed within a relatively short training phase.