| 摘要 |
In speech recognition words to be recognised may be represented by finite state machines and recognition is based on analysing transitions through the machines as an utterance occurs. One value which is required for each state of each machine is minimum cumulative distance; that is the smallest value on reaching one of the states from a starting position, considering all possible paths. Since words are spoken one after another, a finite state machine representing one word has transitions to another machine representing another word. The network of such transitions is complex and varies between different pairs of words. In the present invention, rather than use such a network, each finite state machine is given a start state (SD) at the beginning and an end state (ED) at the end. A Viterbi engine finds the minimum cumulative distance for each normal state of each machine and also determines the minimum cumulative distance for each end state. Traceback pointers for each end state are determined which indicate the number of transitions traversed in reaching that end state. A further distance dependent on the traceback pointer for each end state is added to that state to form a word ending score. The best score is then used to update start states selected on a grammatical basis, other start states being updated with a maximum value. <IMAGE> |
