Abstract
Advances in chord recognition research using machine learning are hampered by two factors: the scarcity of annotated training data, and the limited complexity of the features and models used. Both problems are intertwined, as with few training examples, increasing the complexity of the model would inevitably lead to overfitting. In this paper we develop a way to address the first problem by exploiting chord annotations from online chord databases. We show how such chord annotations, despite being noisy and lacking exact chord onset times, can be put to use both during the recognition and training stage. We note that the ability to exploit this large untapped resource may enable researchers to also address the second problem: with more training data, one may be able to use more complex models without running the same high risk of overfitting.
Acknowledgements
This work is supported by the EPSRC project EP/G056447/1 and by the PASCAL2 European Network of Excellence and also partially by EPSRC grant number EP/E501214/1. We are grateful to Matthias Mauch, Emilio Parrado-Hernandez, and Sandor Szedmak for useful feedback on an earlier draft of this paper and for interesting discussions.
Notes
3Note the slight abuse of notation: we use chords as indices to the vector P ini and matrix P tr.
4The normalization factor
is used to re-normalize
so that
meets the probability criterion
. Similar operations are done for the three methods presented in this subsection.
5Although Jump Alignment is similar to the jump dynamic time warping (jumpDTW) method presented in Fremerey, Müller, and Clausen (Citation2010), it is worth pointing out that the situation we encountered is more difficult than that faced by music score-performance synchronization, where the music sections to be aligned are noise-free, and where clear cues are available in thescore as to where jumps may occur. Furthermore, since the applications of JA and jumpDTW are in different areas, the optimization functions and topologies are different.
6We should point out that our method depends on the availability of line information. However, most online chord databases contain this, such that the JA method is applicable not only to UCSs from the large e-chords database but also beyond it.
7The following five were missing: Love You To,Within You Without You,Wild Honey Pie,Revolution 1,I Want You (She's So Heavy).
8The aligned ground truth annotations are generated by stripping the timing information away from ground truth annotations and re-aligning them using UCSA.
9To make the results in Section 5.2.2 comparable to that in Section 5.2.3, we only used 52 Beatles songs as the expansion set, which is roughly
of the total data size.
10The songs were Bring it on Down,Cigarettes and Alcohol, Don't Look Back in Anger,Morning Glory, and My Big Mouth. We plan to make the annotations available online shortly.