Research On Audio Classification

With the continuous developing of Internet infrastructure, the network bandwidth and storage capacity have outperformed a new level. A variety of multimedia data has been spreading widely. Audio as the main input of human’s auditory system has become more and more diversified, such as streaming music, voice short message, online open courses. Audio diversification has brought new challenges to the hearing world that is how to effectively classify the audio from the increasing amount of data into different collections.This thesis introduces the basic knowledge of audio classification, application scenarios and traditional classification algorithms. On this basis, we propose two kinds of audio classification algorithms with principle and experimental results.The first algorithm uses the statistics of traditional audio features, including linear predictive coding, short-time average zero-crossing rate, short time average energy and spectral flux. Then, the statistical features are extracted from each audio frame. In the classification procedure, we use the Support Vector Machine as the classifier, and RBF kernel is selected as the mapping function between low dimensional features and high dimensional space.The second algorithm uses sparse coding in machine learning field to learn features from raw audio and get an over-complete set of basis vectors. Then, the basis vectors are used to fit the original audio sampling by linear combination, and take the same classifier as mentioned before for classification model training. The experimental results are stable.The experiments of two algorithms have been proved in reality, showing that they can archive faster and better results than state-of-the-art methods in the specific number and content classification task.