INTERMEDIATE FEATURE QUANTITY CALCULATION DEVICE, ACOUSTIC MODEL LEARNING DEVICE, SPEECH RECOGNITION DEVICE, INTERMEDIATE FEATURE QUANTITY CALCULATION METHOD, ACOUSTIC MODEL LEARNING METHOD, SPEECH RECOGNITION METHOD, AND PROGRAM

摘要

PROBLEM TO BE SOLVED: To provide a technique for calculating an intermediate feature quantity to be used to learn an acoustic model for actualizing high-precision speech recognition without reference to the kind of speech data for learning while suppressing a calculation time needed for recognition processing.SOLUTION: An intermediate feature quantity calculation device includes: a kind identification part 110 which determines, from a speech feature quantity, a kind number j' (where j' is an integer satisfying 1≤j'≤J) corresponding to the kind of speech data from which the speech feature quantity is extracted; and a phoneme intermediate feature quantity calculation part 120 which calculates, as a phoneme intermediate feature quantity from the speech feature quantity and the kind number j', a j'-th kind intermediate feature quantity being a feature quantity to be used to calculate phoneme probability distributions p=(p, ..., p) being a distribution of a probability pin which a phoneme to which the speech feature quantity corresponds is a phoneme with a phoneme number m. The phoneme intermediate feature quantity calculation part includes a j-th kind intermediate feature quantity calculation part which uses a neural network for each integer j satisfying 1≤j≤J to calculate a j-th kind intermediate feature quantity from the speech feature quantity extracted from speech data with a kind number j.SELECTED DRAWING: Figure 2