VOICE SEPARATION DEVICE, VOICE SEPARATION METHOD, VOICE SEPARATION PROGRAM, AND VOICE SEPARATION SYSTEM

摘要

A voice separation device (12) of this voice separation system is provided with: a feature amount extraction unit (121) that extracts time-series data of a voice feature amount of mixed voices; a block separation unit (122) that separates the voice feature amount time-series data into blocks having a certain time width; a voice separation neural network (1b) that creates, from the voice feature amount time-series data blocks, time-series data of a mask for each of a plurality of speakers; and a voice restoration unit (123) that restores the voice data of each of the plurality of speakers from the mask time-series data and voice feature amount time-series data of the mixed voices. In the creation of the mask time-series data for each of the plurality of speakers, the voice separation neural network (1b) uses time-series data of a block which temporally precedes the present time, in a forward-direction LSTM neural network, and uses time-series data of a block constructed by a predetermined number of frames which temporally succeed the present time, in a reverse-direction LSTM neural network.