SPEECH RECOGNITION METHOD USING FEATURE COMPENSATION BASED ON DEEP NEURAL NETWORK

摘要

The present invention relates to a speech recognition method by using a feature compensation technique based on a deep neural network (DNN). More specifically, the present invention includes: (1) a step of learning the DNN by using learning data mixed with a noise; (2) a step of calculating a compensated feature vector or a posterior state probability by applying test data to the learned DNN; and (3) a step of drawing a speech recognition result by using the compensated feature vector or the posterior state probability calculated in step (2). According to the speech recognition method by using the feature compensation technique based on the DNN suggested by the present invention, it is possible to automatically learn the relationship between the noise and a clean speech by learning the DNN by using the learning data mixed with the noise and perform the accurate feature compensation with any of the noises, thereby obtaining the improved speech recognition result by drawing the compensated feature vector or the posterior state probability by applying the test data to the learned DNN and by drawing the speech recognition result by using the same. In addition, the present invention is capable of being applied to speech recognition in various modes according to decoder types in that the present invention is capable of learning the DNN in terms of the feature vector or the posterior state probability and obtaining the compensated feature vector or the posterior state probability via the learned DNN.