面向汉语语音关键词检出的时间集成神经网络研究-计算机科学与技术专业毕业论文.docxVIP

哈尔滨工 哈尔滨工业大学工学硕士学位论文 A bstract A s one key research field and application hotpot of speech recognition, K eyw ord Spotting(K W S) technology has m ade a significant im provem ent in recent years. A lthough H idden M arkov M odel(H M M ) has been m ainstream of speech recognition for years, A rtificial N eural N etw ork(A N N ), w ith its strong discrim inatory ability, low com putation cost, high flexibility, has becom e an efficientsolution to K W S. In this paper, w e have an investigation into a novel strategy on K W S, in w hich a new type of tim e-delayed neural netw ork called Tim e-A ccum ulation N eural N etw ork(TA N N ) is adopted, TA N N w ith tw o steps, tim e accum ulation and fram e accum ulation.TA N N is quite a solution to the problem both faced in tem poralsequence pattern classification and tim e w arping in speech recognition. This paperrefers an innovative netw ork training algorithm based on entropy error function(EE F), supplem ented by increased m om entum item s and variable step-learning and other accelerated netw ork convergence algorithm s. Such training algorithm is notonly through learning atseveral points to m ake it adapt dynam ic process in K W S, but also effectively accelerate the speed of A rtificial N eural N etw orks convergence, enable the netw ork to escape som e of the error surface localm inim a in the learning process. In K W S,the paper uses TA N N as the classifier.T he paper also proposes a m ulti-tem plate joint decision-m aking theory,it com es classifier fusion theory. It has im proved the classification of a single integrated in tim e for the shortage, increases the correct rate.T he K W S technique based on TA N N ,correctrate can reach m ore than 80% ,and also can satisfy the request of real-tim e application, w orthy of furtherresearch and exploration. K eyw ords keyw ord spotting; tim e-accum ulation neural netw ork; entropy errorback-propagation -II- 哈尔 哈尔滨工业大学工学硕士学位论文 目录 摘 要.................................