Rhythmic arrangement from finger-tapping audio signals

TitleRhythmic arrangement from finger-tapping audio signals
Publication TypeMaster Thesis
Year of Publication2016
AuthorsNistal, J.
AbstractWe present a novel study on the behavior of human finger-tapping. This technique can be understood as the casual and rhythmic hitting of objects for the expression of a musical idea, even when it is unconscious or sometimes just for stress relief or because of nervousness. The idea that underlies this project is the connection of spontaneous finger-tapping with human-computer interaction and automatic arrangement of percussion. An application under this functional concept would certainly be a useful tool for the home-studio producer. Our first step was to study the behavior of spontaneous rhythmic expression as well as the needs of inexperienced users in the context of rhythm expression. To this end, we first collected a dataset by recording spontaneous finger-tapping patterns performed by subjects from different music backgrounds. Then, an online survey gathering information about the recording was submitted to the volunteers. Further analysis of the survey answers and several spectro-temporal features extracted from the audio content allowed to infer meaningful information about this behavior. Results of this experiment suggested that there are two clear ways for finger-tapping depending on the music training of the performer. We demonstrate the former hypothesis by conducting a classification task between onsets from both finger-tapping methods. We achieved a 99% of accuracy in recognizing drumming expertise levels (expert vs. nave) by means of using onset-related acoustic features. This suggested that people with percussion training are more concerned about timbre aspects and, thus, they take advantage of this quality of sound to provide differences to each stroke when finger-tapping, as opposed to non-expertise individuals. Secondly, we aimed to convert all the gathered knowledge into a creative tool for arranging drum patterns. Therefore, we describe a system for finger-tapping transcription as an underlying step in the usage of this behavior as a mean for improving humancomputer interaction in the context of computer music creation. The system can be divided into three parts: an onset detection and feature extraction step, in which a set of frame-wise time-dependent features are calculated. These features are fed into a k-Means clustering/classification step, in which the feature representation of the finger-tapped onsets are clustered, assigned to a drum sound class and then translated into a drum MIDI-like symbolic representation.