SOUND MODELING AND VOICE RECOGNIZING METHOD FOR LONG-DISTANCE VOICE RECOGNITION BASED ON MULTI-INPUT DEEP NEURAL NETWORK

摘要

To solve the deterioration of recognition performance caused by reverberations in regard to long-distance voice recognition, the present invention provides a deep neural network having multi-segment input for processing a reverberation component changing depending on various environments including a spatial space, a microphone distance, and the like, and a distance-independent voice modeling method using the same. According to the present invention, for long-distance voice recognition, segments having frames of various lengths are formed to form a plurality of multi-input segments, and convolution matrixes having different sizes are applied to each of the segments to obtain output having the same capacity for each of the segments, and then, if convolution matrixes having the same size are outputted, convolution matrixes of all of the segments are compared to obtain a maximum value for building a sound model.;COPYRIGHT KIPO 2018