| Summary: | 碩士 === 國立成功大學 === 電機工程學系碩博士班 === 94 === Different kinds of sound have different properties in our life environment, and we can make out surroundings by recognizing and understanding these properties of environmental sounds. For example, when we hear the fire alarm sound, we can judge there must be fire happening. It will be a great help to us for monitoring surrounding environment if we can classify and identify in accordance with the sound information, especially for the deaf person and security system. Besides, as mentioned in the former article, large amount of information is recorded in files format of audio. Making use of audio classification will be contributive to us for searching the audio segment we want.
In this thesis, we present a home environmental audio classifier based on support vector machine (SVM) and independent component analysis. We use independent component analysis to extract the audio feature. This technique can extract independent components based on statistical characteristics. The proposed audio features can be categorized as three sets. The first feature set is perceptual features which include total spectrum power, subband powers, brightness, bandwidth and pitch. The second feature set consists of MFCC and delta MFCC. The third feature set is the ICA-transformed MFCC feature. This is achieved by transforming the MFCC feature using ICA transform. The ICA transform is literately obtained based on all the training audio data. The audio classifier is designed using probabilistic SVMs. We collect an audio database contained 649 wav files of 15 classes. Experiments demonstrate the proposed sound classifier can achieve a 97.52% classification rate.
|
|---|
