Inverse harmonic clustering for multi-pitch estimation: an optimal transport approach

TL;DR

This paper introduces a novel multi-pitch estimation method using optimal transport theory that improves robustness and performance over existing techniques, especially in noisy and inharmonic conditions.

Contribution

It presents a new framework for multi-pitch estimation that decouples regularization from dictionary design, enhancing robustness and mitigating coherency issues.

Findings

Achieves better estimation accuracy than traditional statistical methods.

Performs comparably or better than network-based methods with less training data.

Demonstrates robustness to inharmonicity and noise in synthetic and real data.

Abstract

In this work, we consider the problem of multi-pitch estimation, i.e., identifying super-imposed truncated harmonic series from noisy measurements. We phrase this as recovering a harmonically-structured measure on the unit circle, where the structure is enforced using regularizers based on optimal transport theory. In the resulting framework, a signal's spectral content is simultaneously inferred and assigned, or transported, to a small set of harmonic series defined by their corresponding fundamental frequencies. In contrast to existing methods from the compressed sensing paradigm, the proposed framework decouples regularization and dictionary design and mitigates coherency problems. As a direct consequence, this also introduces robustness to the phenomenon of inharmonicity. From this framework, we derive two estimation methods, one for stochastic and one for deterministic signals, and propose efficient numerical algorithms implementing them. In numerical studies on both synthetic and real data, the proposed methods are shown to achieve better estimation performance as compared to other methods from statistical signal processing literature. Furthermore, they perform comparably or better than network-based methods, except when the latter are specially trained on the data-type considered and are given access to considerably more data during inference.