Subjective evaluation of music similarity system based on onsets

We have proposed a music recommendation system based on the time difference between adjacent onsets previously. The system also employed the rough longest common subsequence (RLCS) algorithm for similarity measurement. In this paper, we report our results for subjective (listening) comparison of our system with the Musly system. The results show that, two out of five query songs, the audiences felt that the recommended clips from the proposed system are more similar to queries. Since our system uses only “onsets” for similarity measure, as opposed to both temporal and spectral features in other systems, to determine the similarity of audio clips the subjective results confirm that onset information is an important factor to determine the similarity of two audio clips. Thus, a music recommendation system might include the onset features for computing similarity scores.