Driftin' Down the Scale: Dynamic Time Warping in the Presence of Pitch Drift and Transpositions
Simon Waloschek, Aristotelis Hadjakos · Zenodo (CERN European Organization for Nuclear Research) · 2018
Recordings of a cappella music often exhibit significant pitch drift. This drift may accumulate over time to a total transposition of several semitones, which renders the canonical 2-dimensional Dynamic Time Warping (DTW) useless. We propose Transposition-Aware Dynamic Time Warping (TA-DTW), an approach that introduces a 3rd dimension to DTW. Steps in this dimension represent changes in transposition. Paired with suitable input features, TA-DTW computes an optimal alignment path between a symbolic score and a corresponding audio recording in the presence of pitch drift or arbitrary transpositions.