Having worked on genre theories for quite a while, and having faced the problems emerging from intra-disciplinary approaches that often seem to be scarcely attentive to the real content of research in other disciplines, it occurred to me that something radically new was needed: not just an interdisciplinary study, but a study aimed at creating bridges among disciplines, in order to build common knowledge, and a shared language to describe genre-related phenomena. I will finish my overview by illustrating some details of a research project.
The project's aim is to combine the perspectives on music categorization of many and sometimes 'distant' disciplines, like musicology, popular music studies, ethnomusicology and music anthropology, semiotics, sociology and computer science, capitalizing on previous research, and learning from its successful results and from its flaws. Nothing similar has been attempted to date. What is needed is a comprehensive interdisciplinary model for the understanding of music categorization, with a degree of formalization that will allow computer scientists to translate it into flow diagrams and programs, while respecting the complexity of phenomena taking place in many different musical and societal contexts. In the language of computer science this would be defined as the definition of an ontology, that is, a full description (naming, typology, properties) of the conceptualizations of phenomena and their interrelations in a domain of discourse (Gruber 1993). From an anthropological and sociological perspective, the target would be a 'thick description' (Geertz 1973) of music events, allowing the understanding of music practices in the framework of their 'living' context. From a semiotic and anthropological perspective, the study would focus on conventions agreed upon within communities (Morgan 2014), and the musical, linguistic, behavioural, proxemic codes, etc., emerging from that work (Fabbri 2007a, 2007b). For musicology, the challenge is to relate categorizing processes and naming conventions to features of musical structure, and to their historical and social meaning.
If successful, the project would have a considerable impact both on scholarship in the disciplines involved, and on industrial sectors and media, for which music categorization is vital: the live music industry, the phonogram industry, radio and television, Internet. While a better comprehension of music categorization processes is expected to lead to a further integration of musicological disciplines, it could also improve greatly the precision of music recommendation systems, offering important new resources to web-based music applications.
Categorization theories
In the first phase of the research project, different approaches to music categorization should be compared. As I have shown, the debate about categorization processes performed by humans in classifying music is still an open issue. Two main theories can be distinguished, the first one based on conventions, and the second based on prototypes; some studies (Guaus 2009) debate about whether the categorization process is a 'positive' attitude (i.e. one is aware of some given rules or prototypes and decides whether a given music event belongs to a category or not) or 'negative' (i.e. one is not sure to which category a given music event does belong, but is sure about which categories it doesn't belong to).
Genre descriptors
A genre descriptor is any information that may help in the categorization of music. At first glance, two main families of descriptors can be observed: descriptors related with different facets of music, that is, descriptors related to melody, harmony, rhythm or timbre, or other 'intra-musical' parameters, and descriptors derived from the social implications of music (that is, 'para-musical' factors), like, fashion, age, economic interests, and use in social networks, among others (Sordo 2012). Not all musical genres are characterized by the same type of descriptors. Some genres may be characterized by some musical features, while others are determined by the calendar (e.g. Christmas music in Western cultures), or other industrial criteria (Pachet and Cazaly 2000). It must be considered that a descriptor that can be crucial for one specific category can be considered as 'noise' for other categories. Including all descriptors in all categorization processes can lead to a decrease in the accuracy of the model, unless proper weighing factors are introduced.
Time evolution
As I have discussed above, categories in music are in constant evolution. Many musical genres emerge every year (for example in so-called 'electronic music'), some others die. As has occurred since early times, music's evolution has been directly related to technologic development. So, in a given historical period, some contemporary music categories evolved in parallel paths. Even the creation of music genres can be related to specific technology, so specific music descriptors derived from that may not represent a musical category but a musical period.
The goal is to connect diachronic development in music categorization to music histories, in order to build a model of how genres work through time. How do changes in music practices and technologies, as well as in the media landscape, affect categorizing processes? How can categorization be understood in its wider cultural and historical context?
To answer these questions, further investigations into past categorization are needed. Sources can include music publications from the past, including magazines, music press, popular press, newspapers; TV and radio broadcasts; critical essays and music books; advertisements and press releases; etc. Enquiries must encompass different music cultures and their specific practices, in order to challenge any simplistic Anglo-American-centric perspective. Also, folksonomies and labelling practices (Lamere 2008) – which have developed strongly thanks to the Internet over the past twenty years – must be addressed from the point of view of their diachronic and diatopic developments.
'Borders'
If one asks performers/composers which genre labels could be applied to their activity, probably they will start by presenting similarities to one or another strict convention (e.g. musical genre) or prototype (e.g. an artist name, or an existing work). In general, only few musicians agree that they belong to one of the established music genres. Topographic metaphors are often used in describing genres, and since such metaphors are two-dimensional (while genres aren't), 'borders' have to be described as 'fuzzy'; most musicians would probably like to be geo-located close to these fuzzy 'borders' between three or four 'countries'. There appears to be a contradiction between the attitudes of various musical communities (for example, musicians versus listeners/fans) towards the strictness of genre 'rules'.
Knowledge
In machine learning and statistical analysis areas of knowledge, different strategies and algorithms can be applied. In the MIR community there is a long tradition for organizing contests to evaluate the results of different algorithms in relation to different fixed datasets. Evaluation is based on accuracy (precision, recall, f-measure, etc.) and computing time. Some of the best-rated algorithms are built using, for instance, neural networks, which do not allow an easy comprehension of the model itself. Nevertheless, to my knowledge, there are no initiatives attempting to identify the key parameters that can explain the evolution of musical genres over time.
Datasets
As mentioned earlier, the choice of the dataset is crucial for MIR research: biased datasets bring to biased results. Similarly, all musicological, ethnographic, sociological, surveys (etc.) are subject to errors due to the lack of representativeness of samples, weak methodological control, and so on.
The problem of musical categorization can be approached at different levels of detail. For example, trying to understand the taxonomy proposed by the musical industry (record companies, iTunes, Amazon, etc.), is not the same thing as trying to explain the evolution of electronic (dance) music in Northern Europe during the last five years. Since a high level of detail is important, descriptors derived from audio or social sources must have an equivalently high degree of specificity (Herrera 2002). Moreover, not all music listeners move in the same social circles. Some of them buy CDs, others make use of streaming services provided by different companies, while others only go to concerts. Descriptors derived from social sources may often include naive generalizations and need to be treated with particular care, because some collectives may engage in certain social activities while others do not. A descriptor dataset is biased in its construction.
Labels and tags
Since the age of cassettes, people have been able to create their own musical collections according to specific criteria. In some cases, these homemade compilations or playlists are selected according to non-standardized labels such as 'music for driving' or 'house cleaning' (Sordo, Gouyon and Sarmento 2010). Fixed taxonomies are not always useful to the end user beyond buying music in a physical or online store.
Would it be possible (as suggested by Marino 2013) to make a meta-taxonomy of labels? For example: those based on (prosaic) function/use ('music for doing x by', e.g. elevator music, seduction music, music to impress people by, various types of music for dance, background music, foreground music, film music, concert music, etc.); those based on musician understanding of difference; those defined by user group/population; those based on aesthetic preference/values; those based on cultural habitat; those based on musicological observation; those based on computer analysis.
Cultural and linguistic factors have to be taken into close consideration; until recently, research on social tagging has been focused on English (mostly American English) terms. But genres 'exist' (i.e. are used as concepts) in many cultures, and are labelled with terms belonging to many languages. Are such labels translatable? And, conversely, do the same transnational terms (like 'pop', 'rock', 'opera', 'Lied', etc.) have the same meaning in different cultures?
Categorization and music theory
The part of the study dealing with music theory will face several aspects of technical elements. Musical languages, depending on musical genre or situation, do identify similar technical elements with different terms, revealing ideological orientations (Tagg 2013): a study of those technical elements, their occurrences and their names will lead to the creation of a metalanguage and, possibly, to a reform of musical terminologies. It is also important to investigate how similar elements – melodic, harmonic, timbral, rhythmic, in the mix – 'declined' in different situations and played a part in the creation of different musical genres; these 'musical vocabularies' will need to be analysed.
Another fundamental aspect – never studied before – is the categorization of the broad field of 'classical' music: music genres do exist in euro-classical music too, but they seem to be overwhelmed by the huge amount of genres in popular music. In the classical sphere, musical and para-musical activities varied considerably, depending on the venue and the genre, and audiences do tend, still today, to distinguish between genres in what we now call 'classical' music.
The final interdisciplinary model will result from the integration of models derived from the salient points above, with the aim to overcome differences in disciplinary metalanguages, which are obviously apparent now. The project, as stated earlier in this paper, is aimed at 'building a bridge' between musicology (with related Social Sciences and Humanities disciplines) and computer science: both are expected to benefit greatly from the construction. The proposal of a new tagging protocol for audio files is a feasible outcome of the project: if adopted, it would certainly affect the circulation of music currently disadvantaged by improper tagging (classical music, popular music and traditional music from non-Anglophone countries), and favour the usage of digital resources in music education.
Acknowledgements
I would like to thank warmly Marta García Quiñones (Barcelona), Enric Guaus (Barcelona), Jacopo Tomatis (Turin), Jacopo Conti (Turin), Gabriele Marino (Turin), Goffredo Plastino (Newcastle-upon-Tyne), and Philip Tagg (Huddersfield) for their precious suggestions and corrections. Of course, the responsibility for any remaining mistake or lack of clarity is mine.