Query-by-humming on Internet

We present a query-by-humming system on the Internet. It is fundamentally a content-based music retrieval system based on multidimensional spatial indices for high speed similarity searching. This paper discusses technical goals and approaches that will make it possible to use the system on the Internet. The keywords that indicate our goals are retrieval accuracy, quick response, large database, scalability, flexibility, robustness, and usability. We realize these by using some new technologies such as beat-based music data processing, music data splitting by a sliding-window method, database downsizing based on repetitive structure of songs, feature vectors based on timely transition of tones, or-retrieval with multiple feature information, duplicated registration of songs in doubled/halved tempo for a database, and a high-speed retrieval engine with distributed indices. Finally, we present our query-by-humming system, SoundCompass/Humming, that includes these technologies, using our latest GUI.