To provide a voice recognition device in which a highly precise recognition result is obtained even though acoustic features of recognition objects are close to each other and the appearance frequencies of words are approximately same.
In a first sensitivity information processing section 12, a sensitivity value Qk corresponding to the rhythm features of inputted voice data is selected from a prosodic sensitivity model 121. In a voice recognition section 13, voice recognition of the data is conducted using an acoustic model 131 and a language model 132, a score Swn of a candidate word is led and N word candidates Wn having highest Swn are selected. In a second sensitivity information processing section 14, a sensitivity value Rwn is obtained for each of N candidates Wn using a meaning sensitivity model 141. In a sensitivity condition integrating section 15, a recognition score Twn is computed for the N word candidates Wn while a word candidate score Swn is weighted using the value Qk and Rwn. Then, a recognition result output section 16 outputs a word candidate Wn having a highest score Twn as the recognition result.
JP2002366175 | DEVICE AND METHOD FOR SUPPORTING VOICE COMMUNICATION |
JP2004110214 | PRINTER SYSTEM |
WO/2015/027241 | SYSTEMS AND METHODS FOR PROVIDING AUDIO TO A USER BASED ON GAZE INPUT |
ISOBE TOSHIHIRO
Next Patent: N BEST RETRIEVAL METHOD FOR CONTINUOUS SPEECH RECOGNITION