Audio-based computational stylometry for electronic music

TitleAudio-based computational stylometry for electronic music
Publication TypeMaster Thesis
Year of Publication2014
AuthorsRodríguez-Algarra, F.
AbstractIdentifying artists and their stylistic signatures is a difficult problem, especially when only audio files and not symbolic sources are available. This is the most common situation when dealing with Electronic music, so the application of different constructs and techniques that have been proved useful when studying composers that wrote scores is needed. In addition to that, Electronic music increases the complexity of these problems, as timbre and rhythm tend to get more relevance than pitches, durations and chords, facets traditionally emphasized in musical style analysis. The research presented in this dissertation aims at the exploration of the usage of Music Information Retrieval tools and techniques for the stylistic analysis of Electronic Music. For that purpose we have curately constructed a music collection specially addressed for the above-mentioned problems, containing more than 3000 tracks of 64 different Electronic Music artists. The collection has been analyzed with the help of different software libraries, and the extracted features cover different musical facets such as timbre, rhythm, and tonality aspects, and include different temporal scopes (short-term analysis windows, central tendency and dispersion measures for whole tracks, and section summaries). The extracted features are tested in a traditional Artist Identification task, overperforming the previously reported results in similar studies when training a Support Vector Machines model and evaluating it with a holdout test. We also propose a categorization of the most relevant features for capturing the style of Electronic Music artists, distiguishing between “discriminative” and “descriptive” features. Both types are tested experimentaly, achieving satisfactory results. Finally, a detailed qualitative analysis of the results obtained when considering a small group of artists is performed, demonstrating the potential of the analyses that have been developed.
KeywordsSMC MSc Thesis
intranet