Comparing Pitch Spelling Algorithms

A pitch spelling algorithm predicts the pitch names of the notes in a musical passage when given the onset-time, MIDI note number and possibly the duration and voice of each note. Various versions of the algorithms of LonguetHiggins, Cambouropoulos, Temperley and Sleator, Chew and Chen, and Meredith were run on a corpus containing 195972 notes, equally divided between eight classical and baroque composers. The standard deviation of the accuracies achieved by each algorithm over the eight composers was used as a measure of its style dependence (SD). Meredith’s ps1303 was the most accurate algorithm, spelling 99.43% of the notes correctly (SD = 0.54). The best version of Chew and Chen’s algorithm was the least dependent on style (SD = 0.35) and spelt 99.15% of the notes correctly. A new version of Cambouropoulos’s algorithm, combining features of all three versions described by Cambouropoulos himself, also spelt 99.15% of the notes correctly (SD = 0.47). The best version of Temperley and Sleator’s algorithm spelt 97.79% of the notes correctly, but nearly 70% of its errors were due to a single sudden enharmonic change. LonguetHiggins’s algorithm spelt 98.21% of the notes correctly (SD = 1.79) but only when it processed the music a voice at a time.