摘要 |
For automatic speech recognition, reference utterances (r1a, r1b, ...) are first recorded for different words (1, 2, 3), sequences (rmv1a, ...) of successive reference characteristic vectors which are represented at uniform time intervals are formed therefrom, a single model (m1) for the reference utterances (r1a, r1b, ...) whose components consist of compensation functions (a1x, a1y, a1z) is formed for each word from the resulting sequences (mov1a, ...) of model vectors for the reference utterances, a word (ua) to be recognised is processed into a sequence (mv) of characteristic vectors which are represented at the same time interval, and the resulting sequence (av) of representation vectors is compared during comparison steps (v1, v2, v3) to the stored models (m1, m2, m3). |