| 摘要 |
Individual words are recognised in a continuous word chain in similar fashion to a word spoken in isolation, through comparison following non-linear time adaptation in which a sequence of combinations in each case of a test signal derived from the speech signal and a reference signal is sought, for which sequence the summed differences produce a minimum. Since the word boundaries in a continuous word chain are initially unknown, the speech signal must be compared repeatedly with the reference samples of all words of the predefined vocabulary. The reference samples in each case of one word comprise a sequence of individual reference signals. In each case at the end of a sequence of reference signals, specific transition rules are applied for the transition to a new sequence of reference signals. Furthermore, backward pointers are stored which, on the basis of the sequence with the smallest difference sum at the end of the speech signal enable a backtracking of the optimum time adaptation path. The memory space requirement is linearly dependent on the number of words of the vocabulary and the permitted or predefined length of the speech signal. The arrangement for carrying out the method which is also described essentially comprises only these memories and a small number of additional elements, in particular a comparator and a number of intermediate registers, and is thus of simple design. <IMAGE> |
