SIGNAL PROCESSING DEVICE USING NEURAL NETWORK, SIGNAL PROCESSING METHOD USING NEURAL NETWORK, AND SIGNAL PROCESSING PROGRAM

摘要

This signal processing device (10), which estimates a signal in which a reverberation is reduced from an observation signal that contains the reverberation observed from one or more microphones, includes: a power estimation unit (12) which estimates an estimation value of a feature amount corresponding to the power of the signal, in which the reverberation corresponding to the observation signal is reduced, by inputting an observation feature amount corresponding to the observation signal to a neural network that receives a feature amount of a reverberation-containing signal and is learnt to output an estimation value of the feature amount corresponding to the power of the signal in which the reverberation in the signal is reduced; and a regression coefficient estimation unit (13) which estimates regression coefficients in an autoregressive process that generates the observation signal, by using the estimation value of the feature amount corresponding to power that is an estimation effect of the power estimation unit (12).