Method and Apparatus for deep learning based algorithm for speech intelligibility prediction of vocoders

摘要

The present invention relates to a technique evaluating clarity of a voice passing a vocoder and, more specifically, to a method for evaluating clarity of a voice passing a vocoder based on deep learning and an apparatus thereof capable of determining a difference in clarity between an original voice before transmission to a vocoder and a voice passing a vocoder after transmission on a voice transmitted to various kinds of vocoders. The method comprises the following steps. (a) An evaluation module receives an arbitrary original voice and a vocoder passing voice generated by a vocoder. (b) The evaluation module divides the arbitrary original voice and the vocoder passing voice into frames of time units and extracts speech features from each frame to use the extracted speech features as feature vectors. (c) The evaluation module applies the feature vectors to a deep neural network (DNN) regression model to calculate a speech clarity difference for each frame. (d) The evaluation module sums the difference in speech clarity of each frame and calculates a clarity difference score on the entire original voice.