METHOD OF RECOGNISING AT LEAST ONE DEFINED PATTERN MODELLED USING HIDDEN MARKOV MODELS IN A TIME-VARIABLE TEST SIGNAL ON WHICH AT LEAST ONE INTERFERENCE SIGNAL IS SUPERIMPOSED

摘要

PCT No. PCT/DE96/00253 Sec. 371 Date Sep. 8, 1997 Sec. 102(e) Date Sep. 8, 1997 PCT Filed Feb. 19, 1996 PCT Pub. No. WO96/27871 PCT Pub. Date Sep. 12, 1996A special method recognizes patterns in measurement signals. Speech signals or signals emitted by character recognition apparatuses are thereby meant. For the execution of the invention, the hidden Markov models with which the patterns to be recognized are modeled are expanded by a special state that comprises no emission probability and transition probability. In this way, the temporal position of the sought pattern becomes completely irrelevant for its probability of production. Furthermore, the method offers the advantage that new and unexpected disturbances can also be absorbed without the model's having to be trained on them. In contrast to standard methods, no training on background models need be carried out. However, this means a higher expense during the recognition of the patterns, since the individual paths of the Viterbi algorithm have to be normed to the current accumulated probabilities in the path with respect to their probabilities, in order to be able to compare them. The inventive method offers the advantage that only the time segment of the measurement signal also containing the pattern has to be analyzed. An increased probability of a hit is thereby reconciled with a lower computing expense.