Abstract
This thesis investigates how musicologists can use digital audio content analysis (ACA) methods for analyzing large amounts of music. Abundances of music have been digitized in recent years, and the field of Music Information Retrieval has created ACA methods to automatically calculate the so-called features from audio files, such as tempo or tonality. However, the vast majority of musicologists do not yet use ACA tools for analyses, although they can be a means to deal with these abundances of audio files.
I illuminate my research question by seeking inspiration in the broader field of the Digital Humanities for discussing how to apply quantitative techniques within a qualitative field. I examine ACA methods’ practical and epistemological value for musicologists in three case studies: In case 1, I examine machine-learned metrics that allows the creation of features that estimate intuitive qualities of the music. In Case 2, I discuss the epistemological value of an existing big data music analysis. And in Case 3, I apply ACA methods for assisting an analysis of 89 DJ sets played at a recent festival.
Data can be a means to “enhance the perception”. I.e. to observe things that otherwise would have been hard to find. Musicologists can apply data techniques for posing new questions, for example about many pieces of music, or about musical aspects adhering to a subset of a corpus. They can apply them for a “second” empirical opinion on the corpus. With modern software, digital methods allow swift exploration of musical aspects in large corpora for example through visualization, such as to create crude maps of the music. Statistics enable us to generalize, find larger patterns, but also to nuance and find individual differences.
Music analytically, ACA features can both represent new and traditional ways of measuring music. However, there seems to be a current ceiling of about 70-85% correctness for ACA tasks, and this affects the conclusions for large-scale analysis. There is also a current tendency that timbral aspects play a more prominent role than previously. However, the features are created by opaque and complicated algorithms. This impedes the translation from feature to music analysis, and the analysis becomes imprecise or uncertain on a qualitative level, despite exact methods. If musicologists want to exploit the potentials of ACA methods, they will have to get used to not understanding the connection between features and music entirely.
I illuminate my research question by seeking inspiration in the broader field of the Digital Humanities for discussing how to apply quantitative techniques within a qualitative field. I examine ACA methods’ practical and epistemological value for musicologists in three case studies: In case 1, I examine machine-learned metrics that allows the creation of features that estimate intuitive qualities of the music. In Case 2, I discuss the epistemological value of an existing big data music analysis. And in Case 3, I apply ACA methods for assisting an analysis of 89 DJ sets played at a recent festival.
Data can be a means to “enhance the perception”. I.e. to observe things that otherwise would have been hard to find. Musicologists can apply data techniques for posing new questions, for example about many pieces of music, or about musical aspects adhering to a subset of a corpus. They can apply them for a “second” empirical opinion on the corpus. With modern software, digital methods allow swift exploration of musical aspects in large corpora for example through visualization, such as to create crude maps of the music. Statistics enable us to generalize, find larger patterns, but also to nuance and find individual differences.
Music analytically, ACA features can both represent new and traditional ways of measuring music. However, there seems to be a current ceiling of about 70-85% correctness for ACA tasks, and this affects the conclusions for large-scale analysis. There is also a current tendency that timbral aspects play a more prominent role than previously. However, the features are created by opaque and complicated algorithms. This impedes the translation from feature to music analysis, and the analysis becomes imprecise or uncertain on a qualitative level, despite exact methods. If musicologists want to exploit the potentials of ACA methods, they will have to get used to not understanding the connection between features and music entirely.
| Originalsprog | Engelsk |
|---|
| Udgivelsessted | København |
|---|---|
| Forlag | Det Humanistiske Fakultet, Københavns Universitet |
| Antal sider | 225 |
| Status | Udgivet - 27 okt. 2017 |
Emneord
- Det Humanistiske Fakultet
- Music
- Data
- Digital Humanities
- Digital musicology
- Music Information Retrieval
- Audio Analysis
- Information Science
- Data science
- Musicology
- Music research
- MFCC's
- Large-scale analysis
- Popular music studies