基于前馈型人工神经网络的语音和音乐识别-计算机软件与理论专业论文.docx

Speech Speech and MusiC CIassification Based on Feed—Forward ANN Major：Software and Theory of Computer Name：LiU Q5aohui Supervisor：Associate Professor Ou Guiwen Abstract In thiS paper，the approaches，which based on Feed—Forward Artificial Neural Networks，to speech and music classification in continuous audio are deeply studied．The work consists of two aspects mainly：1．What kind of audio feature vectors can be the basi S of classifieation and which one is the best：2．The creating，training，simulating of Feed—Forward Artificial Neural Networks，and the realizing of speech and music classification aecording the audi0 feature vectors． The main contents of thiS paper including： 1．The corpus was recorded from the radio on the Internet．It consists of three audio documents：pure speech、pure music、speech+music，with 225s each one． 2．The effect of every parameter on the classification are studied．The result of experiments shows that MFCC performed the best． 3．The features of Feed—Forward Artificial Neural Networks are introduced．How to creat，train and simulate the Feed～Forward Artificial Neural Networks and to classify the audio documents in MATLAB 6．5 are descr ibed． 4．Training the network using the Levenberg—Marquardt Algorithm is purposed．A val idat J on set is established in order to indicate the end point of training and to avoid endless training which would make the II network network adap L the training set excess and has weak abil ity to extend to other data． 5．Part]y codes of the MATI．AB funct ions are given for referenee． 6．Data and charts are g iven in this paper．The advantage and disadvantage o[。the methods mentioned in this paper is analyzed． The resu]t shows：to speech，music，speech+music documents， Feed—Forward Artificial Neural Networks perfarm terrific，the average classjfi cation accuracy iS up to 94％． Key words：Speech：Music：Classi Fication：Feed—Forward Artj ficial Neural Network IU 基于前馈型人：I一神经网络的语音；fl_】音乐识别第1章引 基于前馈型人：I一神经网络的语音；fl_】音乐识别 第1章引 言 1．1研究背景 随着电台