摘要 |
PROBLEM TO BE SOLVED: To solve a problem that when a keyword is not included in a speech, the optimum path of the keyword obtained by a viterbi method does not match to to the speech and a cumulative distance eventually becomes small, and then a keyword which is not spoken may easily be outputted as recognition results. SOLUTION: A word spotting speech recognition apparatus has a feature parameter generation part (4) which generates a feature parameter of an inputted speech from the speech, a sound model storage part (5) which stores feature parameters of speeches in unit of sub-word, a keyword model generation part (7) which generates a keyword model from reading data of the keyword outputted from the keyword storage part (6) and the feature parameter outputted from the sound model storage part (5), a keyword distance calculation part (10) which calculates the distance between the feature parameter of the speech and the feature parameter of the keyword, a keyword viterbi calculation part (12) which calculates a keyword cumulative distance while outputting state transition information, and a continuance control part (14) which adds a specified value to the keyword cumulative distance when a self-transition frequency exceeds a self-transition frequency threshold. COPYRIGHT: (C)2004,JPO
|