Bridging Printed Music and Audio Through Alignment Using a Mid-level Score Representation.

We present a system that utilizes a mid-level score representation for aligning printed music to its audio rendition. The mid-level representation is designed to capture an approximation to the musical events present in the printed score. It consists of a template based note detection frontend that seeks to detect notes without regard to musical duration, accidentals or the key signature. The presented method is designed for the commonly used grand staff and the approach is extendable to other types of scores. The image processing consists of page segmentation into lines followed by multiple stages that optimally orient the lines and establish a reference grid to be used in the note identification stage. Both the audio and the printed score are converted into compatible frequency representations. Alignment is performed using dynamic time warping with a specially designed distance measure. The insufficient pitch resolution due to the reductive nature of the mid-level representation is compensated by this pitch tolerant distance measure. Evaluation is carried out at the beat level using annotated scores and audio. The results demonstrate that the approach provides an efficient and practical alternative to methods that rely on symbolic MIDI-like information through OMR methods for alignment.