Sound compassTM - a fast query-by-humming system using multi-dimensional feature vectors

Retrieval using sound infonnation as queries for a music database is intuitive and very useful. However, it is difficult to use hummed tunes as queries because they are often unclear. Thus, the SoundCompass™ music retrieval system employs a similarity retrieval technique to overcome this problem. The retrieval result is a ranked list of songs that are similar to the hummed tune. The most significant ways in which our system are superior to other query-by­ humming systems are that 1) musical data is processed based on "beats" instead of "notes", and 2) the retrieval is done through the use of an index based on multi-dimensional feature vectors. These features allow the system to retrieve songs quickly and precisely even if erroneously hummed tunes are used as queries. The database currently holds over 10,000 songs, and the retrieval time is about one second. The system is able to recognize a song and rank it within the first five places on the retrieval list for about 70% of hummed tunes that are recognizable to human subjects as being a part of the song.