YQX+: Benchmarking Probabilistic Modeling of Piano Performance Expression
Abstract A written music score specifies what notes should be played, but much of the musical meaning of a performance comes from how those notes are interpreted: pianists shape timing, dynamics, tempo, and articulation to communicate structure, style, and emotion. Modeling this expressive aspect of music matters for music education, digital performance rendering, and creative AI systems, and is difficult precisely because the same score admits many different but musically valid interpretations.In this paper, we present YQX+, a benchmarking framework for probabilistic modeling of expressive piano performance. YQX+ shifts the focus from demonstrating a single rendering system to comparing how different models capture the expressive variation observed in human performance. To make that comparison possible, it fixes the prediction task, a shared multi-scale contextual feature representation, and a common evaluation protocol, so that model families are trained and compared on equal footing. Given a symbolic score, every model in the suite predicts the same note-level expressive parameters, including beat period, velocity, timing deviation, and articulation ratio, from features that capture local note context, harmonic texture, phrase structure, and score markings. The suite takes the original YQX Gaussian mixture model (GMM) as its reference point and extends it to XGBoost, variational autoencoders, and conditional flow-matching models.To support perceptual evaluation, we also introduce PerformanceMOS, a data-driven estimator of perceived expressive quality and performance skill from symbolic piano performances. Experiments show that larger-capacity XGBoost and \((\beta)\)-VAE models achieve lower timing and articulation errors than prior baselines, including DExter and VirtuosoNet, and obtain the highest predicted mean opinion scores (MOS). These findings establish YQX+ as a systematic benchmark for probabilistic modeling of piano performance expression, with PerformanceMOS providing a scalable proxy for human listening evaluation. The code, trained models, and evaluation scripts are available at \texttt{https://github.com/kkwsts/YQX-Plus}.
Authors
- Aidan O. T. Hogg (ORCID: https://orcid.org/0000-0001-5501-7799)
- 王皓晨
- Simon Dixon (ORCID: https://orcid.org/0000-0002-6098-481X)
- Jinwen Zhou
- Huan Zhang
- Jinhua Liang
- Yuncong Xie
Institutions
- Queen Mary University of London (GB)
- New York University (US)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11241882/v1
- Primary Topic
- Music Technology and Sound Studies
- Type
- preprint