[Back]

P. Knees:
"Information Retrieval and Recommender Systems for Music Listening and Creation";
Technische Universität Wien / Fakultät für Informatik, 2020.

Advances in Informatics are a driving force in the development of today´s music ecosystem. Apart from efficient compression and streaming technology, Artificial Intelligence-and in particular, machine learning-has paved the way for new applications for music consumption, as well as for music creation, and enabled users to access sound and music in unprecedented ways. On the consumption side, tools and interfaces such as recommender systems, automatic radio stations, or active listening applications allow users to navigate the virtually endless spaces of music repositories. On the creation side, software tools like digital audio workstations, new musical instruments, and sound and sample browsers support creativity.
In this thesis, I highlight my contributions to the fields of information retrieval and recommender systems with applications in music listening and creation. Beginning with music listening, I position my work and present methodology, goals, and challenges by focusing on music similarity metrics based on explicit and implicit contextual metadata and approaches to music recommendation.
I further detail recently developed machine learning based approaches to music content analysis, enabling applications in music listening as well as creation. Focus is given to the specific case of drum transcription, in particular by training recurrent and convolutional neural networks. A crucial aspect of this and related work is the generation and acquisition of ground truth data, both for training and evaluation.
For the purpose of music creation, the output of drum transcription can serve as a basis for creativity supporting tools and co-creation of humans and AI, for instance by training generative models. As an example for the methodological steps and challenges involved, the task of automatic drum pattern variation is investigated further. Again, the question of evaluation plays a crucial role in these scenarios, posing a central challenge for co-creation systems and generative AI in general. In addition to the increased need for user feedback in system evaluation, methodologically, this requires overcoming purely system-centric evaluation paradigms for future AI systems. In this light, I discuss my work on user-centric development of information retrieval and recommender systems for music creators and point at challenges ahead.