Pitch Spelling

Pitch Spelling
Estimation of note names from their pitch in number of semitones.

We revisit the old problem of Pitch Spelling (PS), and local and global tonality guessing, with a new algorithm for their joint estimation. It searches for an optimal spelling based on current conventions for music notation, and musicological tools such as distances between tonalities.

Some evaluations of this algorithm were conducted on a large monophonic dataset and a challenging classical piano dataset, and show a higher degree of accuracy than for current state-of-art system for PS.

Moreover, it has been adapted for processing written jazz music, and evaluation on datasets of jazz lead sheets (Real Book), complex jazz soli transcriptions (Charlie Parker Omnibook), and bass lines have also shown very good results.

This joint work with Augustin Bouquillard has been presented at TENOR 2024 and CMMR 2025. It is implemented in the system PSE.