An Unsupervised Audio Segmentation and Classification Approach

This paper presents an unsupervised audio segmentation and classification approach. First, the multiple change-point segmentation is adopted, and a new feature named Mel-ICA is introduced to improve it. An audio type "uncertain" is proposed to represent mixed type. Three features of each sub-segment are extracted using Fourier and wavelet transform. Then, classification is performed over each sub-segment based on feature threshold, and the majority rule is applied to determine the final type. The experimental results have shown that the false alarm rate decreased using Mel-ICA, and high accuracy of classification achieved.