Exploiting Frequency, Periodicity and Harmonicity Using Advanced Time-Frequency Concentration Techniques for Multipitch Estimation of Choir and Symphony

To advance research on automatic music transcription (AMT), it is important to have labeled datasets with sufficient diversity and complexity that support the creation and evaluation of robust algorithms to deal with issues seen in real-world polyphonic music signals. In this paper, we propose new datasets and investigate signal processing algorithms for multipitch estimation (MPE) in choral and symphony music, which have been seldom considered in AMT research. We observe that MPE in these two types of music is challenging because of not only the high polyphony number, but also the possible imprecision in pitch for notes sung or played by multiple singers or musicians in unison. To improve the robustness of pitch estimation, experiments show that it is beneficial to measure pitch saliency by jointly considering frequency, periodicity and harmonicity information. Moreover, we can improve the localization and stability of pitch by the multi-taper methods and nonlinear time-frequency reassignment techniques such as the Concentration of Time and Frequency (ConceFT) transform. We show that the proposed unsupervised methods to MPE compare favorably with, if not superior to, state-of-the-art supervised methods in various types of music signals from both existing and the newly created datasets.