Speech recognition error analysis apparatus, method, program, and recording medium therefor

摘要

P To specify a part being likely to cause error in recognition in a language model. PSOLUTION: In a method and voice recognition processings are performed for a voice signal by using the language model and recognized word strings are assigned. The erroneously recognized word strings constituted by one or a plurality of continuous words which do not agree with the correct word strings corresponding to the recognized word strings in the recognized word strings and a section of erroneous recognition constituted by the erroneously recognized word strings and each of one word before and after the erroneously recognized word strings are extracted out of the recognized word strings. A set of two erroneous words in a START part constituted by the first word in the section of erroneous recognition and the first word in the erroneously recognized word strings is extracted. A set of two correct words in the START part constituted by the first word in the section of erroneous recognition and the first word in the correct word strings corresponding to the erroneously recognized word strings is extracted. Each of word chain probability of the set of two erroneous words in the START part and the set of two correct words in the START part is computed by using the language model. The set of two correct words in the START part having lower word chain probability than the word chain probability of the set of two erroneous words in the START part is extracted. PCOPYRIGHT: (C)2009 and JPO& INPIT