Research On Audio Classification Under Complex Enviroment

Audio information is one of the most important sources of human perception. It is urgent to organize mass audio files according to the semantic description with the rapid increasing of audio and video files. Audio classification is the key problem, and it has become the research hotspot and has very large theoretical and practical values in Pattern Recognition and Artificial Intelligence domain. This dissertation discusses several key problems around audio classification under complex environment, including sample selection, feature extraction and semi-supervised learning. The main contents of this dissertation can be summarized as follows:1) A training sample for annotation selection algorithm based on clustering is proposed.Sample based method is the most effective way for classifier design, so the quality and quantity of training samples become the most crucial factor which affect the classification performance. To reduce the mannual annotation work, an idea,which most useful audio segments are selected and only these ones are annotated, is discussed. The clustering method is used to mine more useful segments without any supervised information. As a result, the classification precision is improved using the more useful samples with the same annotation work.2) A Gaussian Mixture Modem based Discrimination Maximization (GDM) feature selection method is given.Audio feature is another important factor for classification. Fewer features are hoped to learn the classifier with good generalization ability.In traditional, filter feature selection method assumes sample distribution fit Gaussian model. However, the sample distribution of many classes of audio is very complicated, and the signal Gaussian can not describe this kind of distribution. In the same time, the discriminative features between different classes are different. The feature selection method, which max the mean separability, is affected by classes easily separated largely. In fact, it is more helpful for classifier to improve the classification precision of most easily confused classes. Therefore, a feature selection method Maxed Worst Separable Classes Separability based on GMM is discussed, and distance between different GMM is used to measure the separability of different classes.3) A classification algorithm in Fisher kernel space based on Gaussian component clustering is presented.Besides training sample and feature, classification model is another crucial factor which can affect the classification performance. There are two statistical model including Generative Model and Discriminative Model. GMM and SVM are the respective algorithms of these two models respectively.With the limited training samples, SVM is applied as the classifier in this paper. Since SVM classifier demands features with equal length, the mean and the standard variance of the frame features are used in traditional. But an unavoidable loss of information exists when frame-level features are transformed into statistical clip level features. There is another problem that the results tend to be worse when the clip is shorter than 1 s. Accordingly, Fisher kernel, which is based on generative model, is applied to generate equal-length features; besides that, a method based on Gaussian clustering is examined to avoid the high dimension.4) An unlabeled sample selection method of semi-supervised learning based on confidence and clustering is proposed.To resolve the problem that audio annotation needs heavy workload in complex environment, semi-supervised learning is introduced to audio classification. When using the TSVM classifier, the performance is not always improved with the increasing of unlabeled samples. That implies that arbitrarily unlabeled samples can not always help the semi-supervised learning with the limited labeled samples. To choose the helpful unlabeled samples, an unlabeled sample selection method based on confidence and clustering is examined.