| dc.description.abstract | Automatic Performer Identification from the symbolic representation of music has been a challenging topic in the field of Music Information Retrieval (MIR). This thesis proposes various approaches for modeling and identifying musical instrumentalists, with a specific focus on pianist identification, an exceptionally challenging task that often requires trained/expert musicians. Performers’ continuously modifying important parameters like tempo and dynamics to stress specific notes or ‘shape’ certain passages in the metrically-notated music are what makes them distinctive in their perfor- mances. By comparing a performance to its notated score and the perfor- mance norm (defined as a quasi-performance calculated by taking the aver- age of all the performances of the same piece), a set of note-level expressive features related to timing, dynamics, and articulation are proposed that are capable of capturing an individual performer’s performance traits. To val- idate the utility of these characteristic features, several statistical models are used to model their distributions, followed by a similarity metric that compares the distribution similarity of a candidate pianist with that of the pianists in the dataset. The identification is done considering the distribu- tion of each individual feature as well as a feature fusion technique. Results show that features related to expressive timing and loudness are the most informative about performers’ styles when fused together, followed by note duration. Hierarchical modelling of music can be useful for performer identifica- tion, as it allows to capture the structure and organization of the music. Specifically, Western classical music demonstrates a distinct hierarchical or- ganization of elements (note, beat, measure, phrase level etc.). Utilizing a convolutional neural network (CNN) for learning hierarchical representa- tions of this data is a suitable approach. In this study, a pianist identification model is proposed that employs a multichannel 1D CNN, designed to exploit the hierarchical nature of Western classical music through the utilisation of a beat-specific kernel in the first layer of the CNN, optimised to extract musically salient features. Although the proposed model achieves good pre- cision, it does not incorporate recurrence and, as such, is not aware of the context of the music, which is highly dependent on context.Central to this research is the creation of the Automatically Transcribed Expressive Piano Performance (ATEPP) dataset. This extensive dataset, comprising 11,742 virtuoso piano recordings spanning over 1,007 hours, serves as a valuable resource. It facilitates the study of performer-specific expressiveness and diverse playing styles in Western classical piano music, providing a substantial foundation for further investigation and analysis. Finally, to address the limitation of CNNs, a more complex and musi- cally motivated model is proposed that utilizes Recurrent Neural Networks (RNNs) and a multi-head attention mechanism over different hierarchical levels to incorporate recurrence and attention. This facilitates the learn- ing of both the local and global dependencies of the music structure and expressive performance. Results from experimental evaluations reveal that the suggested method outperforms the baseline models, demonstrating the model’s discriminative power and ability to learn performer-specific styles. In summary, this thesis aims to advance performer identification in sym- bolic music data by uncovering key expressive features, proposing innovative modeling techniques, and introducing a comprehensive dataset. These con- tributions provide valuable insights and tools for the field of Music Informa- tion Retrieval, enhancing our understanding of performer-specific musical styles. | en_US |
