LPC-10e DNN Voiced/Unvoiced Decision Method Using Deep Neural Network for Linear Predictive Coding-10e Vocoder

摘要

The present invention relates to a sound improvement technology of a vocoder. More specifically, the present invention relates to a voiced/unvoiced sound discrimination method which attenuates distortion of a synthetic sound source through more accurate voiced/unvoiced sound discrimination using machine learning of voiced and unvoiced sound labeled for each frame of a sound source. According to the present invention, in the case of synthesized sound through accurately discriminated voiced/unvoiced sound, synthetic noise is significantly reduced and higher sound quality can be achieved in the synthesized sound. The voiced/unvoiced sound discrimination method comprises: a step (a) of generating and inputting voiced/unvoiced sound label information and a discrimination reference value for the voiced/unvoiced sound label information by using a TIMIT database (DB); a step (b) of extracting a first sound feature value, inputting the first sound feature value and the discrimination reference value to a deep neural network (DNN) model, and training a weight and a bias; a step (c) of extracting a second sound feature value, inputting the second sound feature value to the DNN model by a DNN model part, and calculating two output values; and a step (d) of comparing the two output values with the discrimination reference value, and discriminating a random input sound source as the voiced sound or the unvoiced sound.