Speech model value determination method

摘要

The method involves determining speech model values at word passages and considering the probability of the word order during recognition of continuously spoken speech with a larger vocabulary. Speech model values are determined before recognition using test-speech signals. Since not all word sequences actually appear at a reasonable extent of test data, the speech model value for e.g. an M-gram-speech model must be determined from appeared word sequences of M-1 words. These shortened word sequences are counted from each different complete word sequence only once, independent of the actual frequency of the appearance of the complete word sequence. Alternately only shorter word sequences are counted, which appear in the test data exactly once.