Hoppa till huvudinnehåll

Multi-pitch estimation

Some of my recent projects have focused on multi-pitch estimation, i.e., estimating the fundamental frequencies of harmonic sinusoidal structures. Multi-pitch estimation can be used in a variety of applications, one of the most fun ones being automatic music transcription. The picture below shows estimated fundamental frequencies obtained when applying the algorithm PEBSI-Lite to a 15 s excerpt of J.S. Bach's Ach, Lieben Christen performed by a violin, a clarinet, a saxophone, and a bassoon. The blue dots are estimates produced by PEBSI-Lite and the solid lines are the ground truth.

A MATLAB toolbox with an implementation of the PEBSI-Lite multi-pitch estimator as described in An Adaptive Penalty Multi-Pitch Estimator with Self-Regularization may be downloaded here. The open-access version of the paper can be found here.

For signals that are non-stationary, due to, e.g., frequency and amplitude modulation, the time-recursive estimator PEARLS may be used. An implementation of this algorithm may be downloaded here.

In some scenarios, the harmonic structure of the signal is only approximate, i.e., the frequencies of the overtones of each pitch are not perfect integer multiples of the corresponding fundamental frequencies. This is the case for, e.g., the voiced parts of human speech, and the sound produced by stringed musical instruments. In a recent paper, we address this issue by formulating the estimation of the fundamental frequencies as an optimal transport problem.

Sidansvarig:  | 2017-11-13