A Conv-TCN Model for Extracting Dynamic Rhythm Paths in Expressive Piano Performances

Abstract

Accurate characterization of expressive rhythm evolution remains a challenging task because conventional methods primarily rely on static descriptors and fail to capture long-range temporal dependencies. This study proposes a Conv-TCN framework for extracting dynamic rhythm paths from piano performance data by integrating one-dimensional convolutional neural networks with temporal convolutional networks. Temporal features are first extracted from MIDI sequences and encoded to represent local rhythmic structures, while dilated temporal convolutions model long-term dependencies and gradual expressive variations. A multi-scale fusion strategy further combines rhythmic information across different temporal resolutions to generate continuous and interpretable rhythm trajectories. Experimental results demonstrate that the proposed method achieves an MSE of 0.032, smoothness of 0.020, KL divergence of 0.052, and similarity of 0.83 under noisy conditions, outperforming several representative sequence modeling approaches. Beyond music performance analysis, the proposed temporal feature extraction and dynamic trajectory modeling strategy provides a methodological reference for processing complex non-stationary signals, with potential value for electromagnetic signal characterization, antenna measurement data analysis, and propagation-related time-series interpretation.