Polina Kozarovytska
Beat Tracking for Genre-Related Datasets Obtained by Clustering
(direct link to PDF)
Abstract
Beat tracking is a fundamental task in music information retrieval. Its performance is often influenced by genre characteristics, as different musical styles exhibit distinct rhythmic structures. For instance, classical music often shows greater metric complexity than hip-hop tracks. This thesis investigates genre-specific beat and downbeat tracking using the Beat This! model proposed by Foscarin et al. (2024) as a baseline. To this end, a pipeline for grouping tracks into genre-related clusters is introduced. Spectrogram inputs are processed by a pretrained genre classification model, from which encoder embeddings are extracted. These representations are subsequently clustered using a dimensionality reduction technique and an unsupervised clustering algorithm. For each resulting cluster, a dedicated beat tracking model is fine-tuned from the baseline model initially trained on the full dataset. In addition, hyperparameter optimisation is applied both to the baseline and to the cluster-specific models, with separate objectives for beat and downbeat tracking. The proposed approach is evaluated on the GTZAN test set using beat-tracking metrics such as F1 score, CMLt, AMLt, and annotation coverage ratio. The results suggest that more fine-grained clustering solutions lead to better performance compared to coarse partitions. Furthermore, it may be advantageous to perform genre-specific fine-tuning only for clusters that are challenging for the beat tracking model, such as those containing classical music. On the other hand, for clusters on which the model already shows steady performance, fine-tuning should be omitted, as the model trained on the entire dataset may be sufficient. Statistically significant improvements in the F1 and CMLt beat-tracking metrics are achieved by the optimised baseline model. In contrast, none of the proposed models outperform the original baseline for these metrics in downbeat tracking, indicating the difficulty of this task. Overall, the proposed genre-specific beat-tracking system achieves a statistically significant improvement in the AMLt beat tracking score.