Combining audio-based similarity with web-based data to accelerate automatic music playlist generation

• Authors: Peter Knees;Tim Pohle;Markus Schedl;Gerhard Widmer
• Venue: MIR '06 Proceedings of the 8th ACM international workshop on Multimedia information retrieval
• Year: 2006

We present a technique for combining audio signal-based music similarity with web-based musical artist similarity to accelerate the task of automatic playlist generation. We demonstrate the applicability of our proposed method by extending a recently published interface for music players that benefits from intelligent structuring of audio collections. While the original approach involves the calculation of similarities between every pair of songs in a collection,we incorporate web-based data to reduce the number of necessary similarity calculations. More precisely,we exploit artist similarity determined automatically by means of web retrieval to avoid similarity calculation between tracks of dissimilar and/or unrelated artists. We evaluate our acceleration technique on two audio collections with different characteristics. It turns out that the proposed combination of audio-and text-based similarity not only reduces the number of necessary calculations considerably but also yields better results, in terms of musical quality, than the initial approach based on audio data only. Additionally, we conducted a small user study that further confirms the quality of the resulting playlists.