This article brings forward the question of which acoustic features are the most adequate for identifying beats computationally in acoustic music pieces. We consider many different features computed on consecutive short portions of acoustic signal, among which those currently promoted in the literature on beat induction from acoustic signals and several original features, unmentioned in this literature. Evaluation of feature sets regarding their ability to provide reliable cues to the localization of beats is based on a machine learning methodology with a large corpus of beat-annotated music pieces, in audio format, covering distinctive music categories.
Confirming common knowledge, energy is shown to be a very relevant cue to beat induction (especially the temporal variation of energy in various frequency bands, with the special relevance of frequency bands below 500 Hz and above 5 kHz). Some of the new features proposed in this paper are shown to outperform features currently promoted in the literature on beat induction from acoustic signals.We finally hypothesize that modeling beat induction may involve many different, complementary acoustic features and that the process of selecting relevant features should partly depend on acoustic properties of the very signal under consideration.
Identifying the Russian voiceless non-palatalized fricatives /f/, /s/, and /ʃ/ from acoustic cues using machine learning