| 摘要 |
A method and apparatus for automated speech recognition wherein an unknown speech input is divided into words on merit, the words are divided into fixed length leading portions and consequentially variable length remaining portions, both portions being further subdivided into fixed length plural frames, each frame being matched against a fixed set of features, determined to correspond to one of the same and tagged with a label representing that feature, the fixed length label string generated from the leading portion of a word being used to aggregate a probability value for the string by accessing individual probability values of occurrence from a leading portion table of labels against first portions of known words and the remaining portion label string being correspondingly processed against a separate remaining portion table, the two sets of aggregate values being compounded by word portion pairs and the words corresponding to the highest compounded values being selected as candidate words for further processing to produce a result, whereby execution of the first portion processing of a word can be undertaken while the remaining portion of the word is being received and the selection of candidates includes a measure of the time of occurrence of a determined feature in a word. The tables are established in a learning phase in which known words are input, divided, framed, featurised and labelled, the probability values for the labelled features being derived by accumulating time dependent weighting factors for given labels in given word, normalising the totals and smoothing the two sets of results. |
