A comparison of pitch chroma extraction algorithms

Pérez Fernández, Miguel; Kirchhoff, Holger; Serra, Xavier

doi:http://dx.doi.org/10.5281/zenodo.6573083

A comparison of pitch chroma extraction algorithms

License

Citation

Pérez M, Kirchhoff H, Serra X. A comparison of pitch chroma extraction algorithms. In: Michon R, Pottier L, Orlarey Y, editors. Proceedings of the 19th Sound and Music Computing Conference (SMC/JIM/IFC 2022); 2022 Jun 05-12; Saint-Étienne, France. [place unknown]: SMC Network; 2022. p. 222-9. DOI: 10.5281/zenodo.6573083

Abstract

The pitch chroma is a popular way to represent pitch information in an octave independent way, with applications in automatic chord recognition, cover song identification, audio-to-score alignment, and others. Early chroma extraction algorithms employed expert knowledge to derive pitch chromas from short-time spectra. With the rise of deep learning, the emphasis moved from algorithm design to the structure of the network and the selection of appropriate training data. The approaches perform differently for various types of audio input. We conducted a set of experiments in order to explore the qualitative properties that each algorithm exhibits. These include how the number of concurrent pitches influences the chroma representation, and how noise or unpitched percussion can degrade the performance of the algorithms. We performed a quantitative analysis of various algorithms under these scenarios. The results show that chromas based on deep learning show huge potential, especially when it comes to noise reduction and ignoring non-tonal aspects of the music. However, we also found that some deep learning based chromas fail to accurately detect pitches at lower polyphony levels. We reflect on these results and discuss some paths to improvements for future chroma extraction algorithms.