Retrieving Musical Information Based on Rhythm and Pitch Correlations
Will Archer Arentz, Magnus Lie Hetland, Bjørn Harald Olstad · Journal of New Music Research · 2005
Music information retrieval and query-by-humming systems have recently been given much attention. One of the reasons for this is the diversity of attractive applications that can be developed with these techniques. Imagine having a tune on the tip of your tongue, but you cannot really place it. What is the name of the song and who is the artist? The proposed method is designed to help in pursuing such a request. Both pitch and rhythmic information is utilized to determine the most closely matched song for any given theme. This also includes polyphonic songs, which may not contain the wanted theme in the main melody, but rather in one of its accompaniments. Due to humans' poor ability to accurately reproduce a piece of music, whether it is caused by poor memory or poor music skills, such methods must be robust in a way that makes people's perception of similarity highly correlated with the distances calculated by the algorithm. The proposed dynamic programming algorithm finds 93 per cent of the wanted hits among the top 10 results when the timing is distorted with a standard deviation of 0.2 seconds and the pitch is distorted with a standard deviation of 0.3 notes, using 7-note queries and a 1,564-song MIDI database.