Features for comparing tune similarity of songs across different languages

Finding tunes that are similar across languages and cultures offers new ways to study global musical influences and similarities. From a signal processing point of view, we find that the availability of vocal music tracks provides us a means for computing tune similarity even in the presence of language differences. While the different acoustic characteristics of each language add to the inherent ambiguity in these kind of problems, the guarantee that a vocal track exists can be a boon in disguise. For this purpose we use the Multi Band Autocorrelation Peak (MBAP) features, extracted in multiple bands providing complementary information which helps to improve the accuracy. Results obtained on a classification task suggest that these features can outperform traditional features like Chroma which capture information from the entire spectrum. Alignment cost using the dynamic time warping algorithm was used a classification metric on a dataset of songs obtained from Youtube.