An Effective Vocal/Non-vocal Segmentation Approach for Embedded Music Retrieve System on Mobile Phone

With the growing bodies of MP3 songs in Internet, content-based analysis plays an important role for its retrieving and management. Due to most useful information is carried by vocal portions, it is necessary to separate the vocal segments from music. This paper presents a method for vocal/non-vocal segmentation, which uses a new feature extracted directly from MPEG encoded bitstream to avoid the computational cost of completely decoding process. In contrast to conventional classification method based on statistical model, in our method, the similarity matrix is first introduced to partition the input into a series of portions; then the SVM (support vector machine) classifier is employed for vocal/non-vocal classification for each portion, finally, a smoothing method is adopted to correct the misclassification errors brought by the classifier. Experiments show the proposed method not only has lower computational complexity, but also has better accuracy rate of the vocal and non-vocal classification under a broad range of signal noise ratio than the original ones.