Acoustic model learning device, speech recognition device, acoustic model learning method, speech recognition method, acoustic model learning program, and speech recognition program

摘要

PROBLEM TO BE SOLVED: To perform an adaption to the acoustic condition of an acoustic model precisely, and to perform voice recognition with reduced retardation.SOLUTION: A feature quantity extraction unit 11 extracts a first input feature quantity and a second input feature quantity for each frame of speech data. In addition, a bottleneck feature quantity calculation unit 121 calculates a bottleneck feature quantity using a bottleneck feature quantity calculation model on the basis of the second input feature quantity. Further, an average calculation unit 122 calculates an auxiliary feature quantity which is an average of bottleneck feature quantities in a plurality of frames. In addition, an acoustic condition feature quantity calculation unit 131 calculates an acoustic condition feature quantity using an acoustic condition feature quantity calculation model on the basis of the auxiliary feature quantity. Also, an output probability calculation unit 132 corrects an acoustic model parameter on the basis of the acoustic condition feature quantity. Also, a word string searching unit 133 outputs word strings using an output probability of an acoustic model and a generation probability of a language model.SELECTED DRAWING: Figure 13