Multiple Fundamental Frequency Extraction for Mirex 2011

This extended abstract outlines an efficient approach for the extraction of multiple fundamental frequencies from polyphonic musical audio. The algorithm consists of three analysis steps. At first a multi resolution spectral analysis is performed on the audio signal. Then, the most salient pitches are identified using a pitch extraction algorithm, which is designed to identify the predominant pitch in polyphonic audio. Finally, distinct high level tone objects are created and tracked over time: the most salient pitch of the current analysis frame may start a new tone object. All active tone objects are jointly evaluated in order to estimate their pitch and magnitude and establish timbre information. The proposed algorithm is a front-end for a melody extraction system, which places a high priority at the most salient tones and at the processing of a human singing voice. Hence, the evaluation results for the multiple-F0 analysis are also discussed in the scope of melody extraction.