| 摘要 |
<p>A data structure includes data indicating a set of strings of character or phoneme candidate identifiers. Each identifier indicates a machine discriminatable candidate type, and the data are accessible to determine whether a string of candidate identifiers is in the set. To produce the data structure, the set of strings is obtained using a list of words. All possible strings up to a given length are obtained, and those that have a high probability of being one of the words are retained. Also, the words are expanded into strings by using probable candidate identifiers for each character or phoneme. If necessary, the number of candidate types is increased or decreased to obtain a satisfactory set of strings. The data structure can include information relating to each string in the set; for example, it can be a finite state transducer with data units, each including a candidate identifier and a character or phoneme identifier, so that the characters or phonemes of a word can be obtained from it. A processor receives a string of candidate identifiers from a candidate discriminator and searches the data structure for the identifiers that occur in the string. If a search succeeds, data relating to the matched string are obtained, such as a word it is likely to be or an identifier of the string. If the string is likely to be more than one word, one of the words is selected and its characters or ph,one,mes are provided as output. A system can include parallel recognition units, each with a respective data structure. The sets of strings can correspond to words with features such as' part of speech font speaker, a language, or a specialized vocabulary. A results processor can receive the results from the recognition units and use them to obtain the characters or phonemes of a word that is likely to be the string being recognized.</p> |
