High Time-Resolution Estimation of Multiple Fundamental Frequencies

This paper presents a high time-resolution strategy to estimate multiple fundamental frequencies in musical signals. The signal is first divided into overlapping blocks, and a high-resolution estimate made of the short-term spectrum. The resulting spectrum is modified such that only the most relevant spectral components are considered, and an iterative algorithm based on earlier work by Klapuri is used to identify candidate fundamental frequencies. Finally, a context-based rule is used to improve the accuracy of fundamental frequency estimates. The performance of this technique is investigated under both noiseless and noisy conditions, and its accuracy is examined in cases where the polyphony is known and unknown a priori.