Co-training Approach for Label-Minimized Audio Classification

Audio classification is an important preprocess to the audio data. However, lots of manual labeled data are needed for training models. In order to solve this problem, we evaluate a semi-supervised machine learning algorithm called co-training for content-based audio classification. The audio is divided into there classes: pure speech, pure music and speech mixed with music. We consider the audio features as views and minimize the labeled data quantity by using co-training algorithm. The experimental results on the VOA Special English show the effectiveness of the co-training algorithm for audio classification.