Audio Content Classification Method Research Based on Two-step Strategy

Abstract

Audio content classification is an interesting and significant issue. Audio classification technique has two basic parts: audio feature extraction and classifier. In general the audio content classification method is firstly to identify the original audio into text, then use the identified text to classify. But the text recognition rate is not high, some words that good for classification are identified by mistake causing that the classification effect is not ideal. In order to solve these problems above, this paper proposes a new effective audio classification method based on two-step strategy. In the first step the features are extracted by using the improved mutual information and classified with Naïve Bayes classifier. After classification of the first step, an unreliable area is determined, and samples with features in this area go on to be classified with the second step. In the second step, textual features extracted with CHI statistic method are used to build a text feature space model. Then audio features containing MFCC and frame energy are combined together with the text features to build a new feature vector space model. Finally, the new feature vector space model is classified using Support Vector Machine (SVM) classifier. The experiments show that the two-step strategy classification method for audio classification achieves great classification performance with the accuracy rate of 97.2%.