New Metrics For Evaluating The Accuracy Of Fundamental Frequency Estimation Approaches In Musical Signals

This paper demonstrates the importance of assessing the performance of fundamental frequency estimation algorithms on note-level descriptors in addition to frame-level accuracy. Note-level descriptors provide a better description of the human experience of listening to music and thus a more perceptually-relevant evaluation of algorithms than frame-level metrics. This is particularly important for tasks that model human engagement with music, such as the study of expressive music performance. This paper evaluates the magnitude of accuracy differences between the frame- and note-level evaluation metrics through an experiment that compares frame-level accuracy measurements to the accuracy of four note-level frequency-related parameters (perceived pitch, vibrato rate, vibrato depth, and jitter) for two score-informed algorithms. The algorithms' accuracies are compared on multi-track recordings of either vocalists or a combination of violin, saxophone, clarinet, and bassoon both on the original anechoic recordings and versions with artificial reverberation added. The representations performed within the margin of error of each other for frame-level accuracy but had significantly different accuracies on all of the note-level parameters in the anechoic condition. This paper concludes by proposing new evaluation metrics that capture temporal characteristics of fundamental frequency traces.