摘要 |
Method of segmentation of chains of words in the training phase of a connected word recognizer. Of the type that includes the steps of automatic segmentation 2 of the current chain of words to be segmented 1, comparison 3 of the number and identification of the words into which the current word chain being segmented, has been segmented 1 with the values of the number of words and their identification corresponding to the original word chain 4, and decision 5 on the incorporation 8 of the segmentation obtained in Hidden Markov Models when comparison 3 shows that the number of words and their identification are the same. The method is characterized by the fact that when the previous comparison shows that the aforementioned values are not the same, a proportional segmentation 6 of the word chain being segmented 1 is carried out, consisting of dividing this chain into as many segments as there are words in the original word chain 4 and assigning to each segment a duration proportional to the duration of the corresponding word when it is pronounced in isolation 7. <IMAGE> |