Improving score to audio alignment: Percussion alignment and Precise Onset Estimation
Résumé
Score to audio alignment connects events in a score to points on the time axis of the performance
audio signal. By using the score, it facilitates audio indexing. This paper presents some major
improvements, particularly alignment of percussive sounds in polyphonic music in addition to
harmonic notes and precise time-onset estimation.
A detection algorithm for percussive sounds uses correlation with percussive sounds from a
database, to detect template sounds, and finds new references in the performance corresponding to the
template sound looked for. Then a local distance for percussive sound and harmonic notes is used by
a dynamic time warping (DTW) algorithm to perform the alignment of all notes together.
For precise time-onset estimation, different methods based on the energy of the note are
applied on the frames given by the DTW alignment.
To remedy for possible frame errors, an algorithm of tempo control is used to detect the notes which
are found out of the local forward and backward tempi. The algorithm is evaluated on a reference
database including classical, jazz improvisation and popular music. Results show a good reliability
and precise time-onset estimation.