The proliferation of personal computing devices and network connectivity has brought forth a new age of music. Advanced music production software has democratized the creation process, while UGC (User Generated Content) sites like SoundCloud have provided accessible methods for distribution. But good music does not equate to commercial success. Artists who wish to make a profession out of music still need the investment capital and media reach of an increasingly centralized group of music production companies. Traditionally, it’s been the job of the A&R (Artists and Repertoire) division of these labels to find commercially-viable talent from this pool. However, with the number of amateur artists increasing by several orders of magnitude, record labels are at an impasse. How can they maintain their centralized forms of talent acquisition if there are hundreds of thousands of artists to choose from?
To fill this industry gap comes a broad array of data science initiatives. Companies like HITLAB and Musiio have promised to remove the listening of tracks from traditional A&R operations through classification of psychoacoustic features from sound files. Trained on large libraries of previously released tracks, these models can use these psychoacoustic labels to predict commercial viability of an artist’s library of songs without a human listener. Combined with more traditional data analytics methods on social media engagement, these models are used to create “comprehensive” data systems that are claimed to de-risk artist investments and improve the reach of music labels to the growing pool of artists. Warner Media Group has doubled the number of their artist signings by leveraging these tools, and has hired a new Global Head of Data Science to further ‘spearhead the company’s global efforts to make “data-backed business decisions”.
But economic value alone cannot be used as a determining factor in the ethical considerations of this transition. Automated methods for filtering out candidates can have a dramatic effect on the diversity of artists that see investment. Technology that leads to tangible effects to the diversity of artists signed has a direct affect on the political and cultural trends of society as a whole. The use of social engagement analytics and psychoacoustic classification could result in a narrowing of music diversity, forcing musicians to conduct and organize their labor differently to attract investment. Furthermore, the lack of transparency in the data science lifecycle of these programs prevents the public from assessing the severity of these effects.
Psychoacoustics and Prediction
Psychoacoustics is exactly what it sounds like, the study of the psychological effects of sound. Psychoacoustics is not a new field, with soundscape studies detailing correlations between the characteristics of a sound and the emotional response perceived by the human listener. These characteristics can be understood as parameters in a parametric representation of that sound, taking forms like loudness, tonality, sharpness, etc. Companies like Musiio aim to capitalize on this field by analyzing sound files for these psychoacoustic features, tagging these features, and then classifying the song by comparing it to a large library of existing commercially successful music. Debate on the merit of the psychoacoustic features goes beyond my expertise, however questioning the validity of the predictive powers of such an algorithm isn’t. Weights derived from training on an example set of previously successful tracks doesn’t provide insight into what is popular now. It’s representative only of what was popular then. Predicting trends in the characteristics of successful music assumes a static environment (all things equal), however the cultural environment corresponding to songs are ever-changing. A great example is hip-hop, whose prevalence today disrupted a long-standing dominance of more traditional genres. Looking at psychoacoustics alone, a rap song wouldn’t rate commercially successful by an algorithm trained on a library of 80s and 90’s pop and rock chart-toppers. Influenced by the changing cultural environment, audiences began identifying with the hip-hop genre, changing the idea of what a successful song sounds like. Algorithms like these may serve as broad-generalizations that perpetuate established patterns of music, promoting traditionally successful genres over the experimental.
Narrowing Diversity and Centralization of the Music Industry
The use of psychoacoustic classification in A&R filtering processes will inadvertently favor established patterns in music. As a consequence to this, talented artists from underrepresented backgrounds or experimental genres may be overlooked, hindering the industry’s potential for fostering creativity and innovation. If data-driven methods continually favor these norms, a homogenization of music culture will ensue. This lack of diversity could perpetuate existing power structures, stifling the development of new music trends and artistic innovation.
These classification methods may also influence demand-side trends in the music industry. SoundCloud, which has recently acquired Musiio, aims to utilize its classification technology for improving user song recommendations, making it core to SoundCloud’s discovery experience. This is a potential feedback loop, with users listening to songs recommended by an algorithm that optimizes for the same traits as those used for A&R. This would reinforce the idea that certain song characteristics are predictors of hit potential, despite just being the product of the recommendation patterns of the system.
It won’t take long for artists to change their creative methodologies in response to these changes. Songs have already gotten shorter to optimize for compensation structures in streaming services (Wright). Social media analytics have also changed the content of music, with artists focusing on memorable hooks that start immediately to avoid a skip (resulting in a “non-complete heard” label on Spotify that affects rankings). The use of social engagement analytics and psychoacoustic classification will force musicians to organize their labor differently to attract investors. Artists may feel pressured to conform to norms, potentially sacrificing their unique artistic vision.
The Solution: Transparency
Until a causal relationship can be determined between song characteristics and their success, these algorithms serve as potential socioeconomic barriers in the music industry. Understanding when these algorithms are valid shouldn’t be up to companies behind closed doors. With artists ‘livelihoods at stake, record labels must adopt one of the fundamental rules of responsible big data research: auditability. Socially responsible usage of such methods in music would require thorough vetting on the front of predictive performance, equity, and safety. Neither I nor the data science teams at any one firm can verify these traits. Opening up these algorithms to public scrutiny would provide companies with extensive feedback that would get them closer to an equitable solution that serves the public good. Moreover, this transparency would improve public understanding of talent acquisitions at music labels, improving the legitimacy of augmented/automated A&R pipelines.