| dc.description.abstract | The perception of music expression and emotion are greatly influenced by performer's individual interpretation, thus modelling performer's style is important to music understanding, style transfer, music education and characteristic music generation. This Thesis proposes approaches for modelling and identifying musical instrumentalists, using violinist identification as a case study. In violin performance, vibrato and timbre play important roles in players’ emotional expression, and they are key factors of playing style while execution shows great diversity. To validate that these two factors are effective to model violinists, we design and extract note-level vibrato features and timbre features from isolated concerto music notes, then present a violinist identification method based on the similarity of feature distributions, using single feature as well as fused features. The result shows that vibrato features are helpful for the violinist identification, and some timbre features perform better than vibrato features. In addition, the accuracy obtained from fused features is higher than using any single feature. However, apart from performer, the timbre is also determined by musical instruments, recording conditions and other factors. Furthermore, the common scenario for violinist identification is based on short music clips rather than isolated notes. To solve these two problems, we further examine the method using note-level timbre features to recognize violinists from segmented solo music clips, then use it to identify master players from concerto fragments. The results show that the designed features and method work very well for both types of music. Another experiment is conducted to examine the influence of instrument on the features. Results suggest that the selected timbre features can model performers’ individual playing reasonably and objectively, regardless of the instrument they play. Expressive timing is another key factor to reflect individual play styles. This Thesis develops a novel onset time deviation feature, which is used to model and identify master violinists on concerto fragments data. Results show that it performs better than timbre features on the dataset. To generalise the violinist identification method and further improve the result, deep learning methods are proposed and investigated. We present a transfer learning approach for violinist identification from pre-trained music auto-tagging neural networks and singer identification models. We then transfer pre-trained weights and fine-tune the models using violin datasets and finally obtain violinist identification results. We compare our system with state-of-the-art works, which shows that our model outperforms them using our two datasets. | en_US |
