| 摘要 |
<p>In speech recognition, speech values derived from the speech signal are compared with reference values for the individual words of a predetermined vocabulary. A known method of reducing the recognition error rate is to take into account speech patterns, the effect of which is to permit only selected sequences of words. From the theory of formal languages a class of speech patterns known as "context-free grammar", which characterises a relatively flexible speech pattern, is known. To apply this speech pattern in a technical process for recognition of a speech signal, two lists are used, which give the correlation between words and predetermined syntactical categories and the correlation of these categories where necessary with two other categories. Both lists are used for each new speech signal; the system always looks retrospectively to see which category best explains the previous speech segment. At the end of the speech signal, it is possible, on the basis of the category which gives the complete sentence, to trace back to the sequence of the words which gave the smallest total distance sum and which also fits into the speech pattern specified by the two lists.</p> |
