An Approach For Music Melody Exteraction Based On Underdetermined Single-Source Speech Sepatation

With the rapid development of Internet and computer science, the quantity of audio content has an explosive growth. It is more difficult for users to locate multimedia resources. Content-based music retrieval provides users a brand new way to search audio content. It can help user avoid the traditional ways to find out information by text. By using this technology, users can search the music though recording a clip or humming by them. It has a wide application prospect in digital rights management. That technology is a hotspot in content-based information retrieval. Music melody extraction is a key point in content-based music retrieval. But it is still in developing, for some of related tools in blind source separation or in computational auditory scene analysis are not well-developed.This paper proposed a fundamental frequency tracking algorithm by take the power of speech harmonic into account. Based on the character that when human pronounces the vocal cords and track are changing continuous, we hypothesize the correlation of harmonic of nearby speech frame is greater than other frames. And then we use large amounts of data to prove it. After analysis various conditions in multi-fundamental frequency, we design a fundamental frequency method and an algorithm to extract melody from music. The melody extraction algorithm tracks all possible fundamental frequencies according to the harmonic hypothesis. Then discard instrument and other harmonic from the results before to form final melody line. Finally, ADC09 and MIR 1K for MIREX data sets are involved to evaluate. The result shows that the algorithm can successfully extract the main melody. Comparing with the scores of MIREX2011, our method has an acceptable performance and accuracy. Our approach is much simpler than tradition’s, it prevent complex mathematical models, and can also be used in blind separation of multi-speaker or speech noise reduction, etc.