EVS Voice Activity Detection based on Deep Neural Network Using EVS Codec Parameter and Voice Activity Detection Method thereof

摘要

Disclosed are an apparatus for detecting voice based on a deep neural network using an EVS CODEC parameter and a method therefor. The method for detecting voice based on a deep neural network using an EVS CODEC parameter includes the steps of: extracting at least one feature vector used in a voice detecting algorithm of an enhanced voice services (EVS) CODEC from a voice signal for learning; calculating a weight point matrix and a bias vector that are applied to an input layer, a hidden layer and an output layer of the DBN through learning by applying the extracted feature vector to a deep belief network (DBN), which is one of deep neural network (DNN) models; calculating a voice existence probability by using the calculated weight point matrix and the calculated bias vector; and determining whether there is voice by comparing the calculated voice existence probability with an adjustable threshold value. In the present invention, a voice detection performance is improved by using a determination logic newly modeled through the DNN.