Multiple-F0 estimation of piano sounds exploiting spectral structure and temporal evolution

This paper proposes a system for multiple fundamental frequency estimation of piano sounds using pitch candidate selection rules which employ spectral structure and temporal evolution. As a time-frequency representation, the Resonator TimeFrequency Image of the input signal is employed, a noise suppression model is used, and a spectral whitening procedure is performed. In addition, a spectral flux-based onset detector is employed in order to select the steady-state region of the produced sound. In the multiple-F0 estimation stage, tuning and inharmonicity parameters are extracted and a pitch salience function is proposed. Pitch presence tests are performed utilizing information from the spectral structure of pitch candidates, aiming to suppress errors occurring at multiples and sub-multiples of the true pitches. A novel feature for the estimation of harmonically related pitches is proposed, based on the common amplitude modulation assumption. Experiments are performed on the MAPS database using 8784 piano samples of classical, jazz, and random chords with polyphony levels between 1 and 6. The proposed system is computationally inexpensive, being able to perform multiple-F0 estimation experiments in realtime. Experimental results indicate that the proposed system outperforms state-of-the-art approaches for the aforementioned task in a statistically significant manner.