An automatic accompaniment system for the singing voice, which allows the user to begin singing from any point within the piece he/she desires. The system is based on the Dynamic Time Warping (DTW) algorithm, which is used as both a similarity measure as well as an alignment tool between the live performance and the reference. The user’s position within the piece is initially determined with a variation of current Query by humming/singing methods, based on three sets of acoustic features; pitch trajectory, RMS energy envelope and MFCCs. After locating the user, the system adapts the tempo of the accompaniment utilizing a known on-line DTW method. The system is implemented in MATLAB and Java, while audio playback is handled using Max/MSP.
This project was carried out as my master’s thesis in the Music Technology Group of Universitat Pompeu Fabra, supervised by Hendrik Purwins. For more details on the project, check out the SMC 2010 paper or the thesis itself.