Multi-Pitch Estimation of Polyphonic Music Based on Pseudo Two-Dimensional Spectrum

Abstract

Multi-pitch estimation is a fundamental and key problem in music information retrieval, but still remains challenging due to the intrinsic complexity of polyphonic music. To address this problem, a pseudo 2-D spectrum-based method is proposed in this article. The pseudo 2-D spectrum is first constructed to map the time domain signal into the 2-D frequency space, where the harmonic signal exhibits a typical 2-D pattern. Then, pitch estimation is carried out by cross-correlation between the pseudo 2-D spectrum and the fixed 2-D harmonic template. Finally, the pitches of adjacent frames are grouped into pitch contours, where the contours whose lengths are shorter than the minimum note length limitation are discarded. And the remained pitches are refined using the estimates of neighboring frames by removing probable errors and reconstructing estimates. The proposed method exploits the harmonic structure of pitched sounds in a two-dimensional frequency plane, can work in the case where some notes contain few harmonics, and the harmonic overlap proportions are reduced greatly in the harmony cases. The experimental results show that the proposed method achieves promising performance comparing with the state-of-the-art methods on the evaluation datasets, and outperforms the bispectrum-based method on both evaluation datasets.