MusPyExpress: Extending MusPy with Enhanced Expression Text Support

Phillip Long; Hao-Wen Dong; Julian McAuley; Zachary Novack

cs.SDAug 21, 2026

MusPyExpress: Extending MusPy with Enhanced Expression Text Support

Authors: Phillip Long, Hao-Wen Dong, Julian McAuley, Zachary Novack

Organizations: Computer Science & Engineering Department, University of California, San Diego · Department of Performing Arts Technology, University of Michigan, Ann Arbor

Abstract

Current work in modeling symbolic music primarily relies on representations extracted from MIDI-like data. While such formats allow for modeling symbolic music as sequences of notes, they omit the large space of symbolic annotations common in western sheet music broadly known as expression text, such as tempo or dynamics, which specify time- and velocity-dependent controls on the musical composition and performance. To alleviate this gap, we present MusPyExpress, an extension to the popular symbolic music processing library MusPy that enables the extraction of expression text along with symbolic music for downstream modeling. Utilizing this extension, we parse the PDMX dataset to illustrate the wealth of expression text available in MusicXML datasets. Additionally, we introduce multiple generative tasks, including joint expression-note generation, expression-conditioned music generation, and expression tagging, that take advantage of this additional notational information.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of MusPyExpress . Although the format of the base Music object is the same as MusPy’s 6 , we have significantly expanded upon the annotations feature (which exists at both the score- and track-level), adding 28 new annotation types, the most of common of which are shown here. The boolean real_time attribute dictates whether temporal features such as onset or duration are stored in units of time steps (metrical time) or seconds (real time).

Panel of Figure 2 — Figure 2: Expression Text in PDMX . Figure 2(a) shows boxplots of relative expression text density. Throughout this work, “Special Barline” refers to special barline types like double and dotted barlines and excludes the ordinary single barline. We include the relative density of key and time signature changes for reference. Figure 2(b) shows boxplots of expression text usage (i.e. the number of different expression text markings) per song for a select group of composers.

Model	PCE	SC (%)	GC (%)	Perplexity ( $\downarrow$ )
Ground Truth	$2.68\pm 0.00$	$97.40\pm 0.01$	$93.67\pm 0.01$	-
Baseline ( M )	$2.82\pm 0.02$	$93.31\pm 0.45$	$\mathbf{93.87\pm 0.21}$	$2.80$
Joint, Prefix ( M )	$1.75\pm 0.07$	$95.57\pm 0.45$	$97.41\pm 0.19$	2.64
Joint, Anticipation ( M )	$\mathbf{2.72\pm 0.03}$	$92.77\pm 0.46$	$93.93\pm 0.26$	$3.58$
Conditional, Prefix ( M )	$1.83\pm 0.06$	$94.91\pm 0.44$	$94.43\pm 0.32$	$3.07$
Conditional, Anticipation ( M )	$1.58\pm 0.05$	$\mathbf{96.72\pm 0.34}$	$97.80\pm 0.14$	$3.02$

Table 1: Evaluation Results . Objective evaluation results for joint (1) and conditional (2) generation tasks. We evaluate the same metrics proposed in Multitrack Music Transformer (MMT) 5 – pitch class entropy (PCE), scale consistency (SC) and groove consistency (GC) – where a closer value to that of the ground truth is considered better. We also evaluate each model’s perplexity over notes (i.e. ignoring expression text), where lower is better.

Model	Beat	Position	Time	Value	Duration	Total
Baseline ( $\it{M}$ )	11.36	86.34	-	1.26	14.68	0.05
Prefix ( $\it{M}$ )	62.80	93.55	-	55.60	54.15	29.41
Anticipation ( $\it{M}$ )	59.78	92.73	-	50.43	50.11	24.41
Baseline ( $\it{R}$ )	-	-	4.17	13.86	59.02	1.63
Prefix ( $\it{R}$ )	-	-	46.05	57.39	66.54	33.90
Anticipation ( $\it{R}$ )	-	-	40.25	53.32	65.60	27.75

Table 2: Expression Tagging Results . Per-field accuracy of expression tagging (3) models on validation set. As a baseline, we examine a model that simply predicts the most common value of each field. Baseline differences between the metrical and real timing schemes may arise because the real time representation truncates songs at 60 seconds, excluding expression text in longer pieces and thereby altering the distribution of field values.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Expression Text Type	Count
Tempo	1,082,640
Dynamic	1,069,492
Text	467,479
Barline	336,029
Rehearsal Mark	176,579
Hair Pin Spanner	164,014

Appendix

Table 3: Expression Text Frequencies . Frequencies of different expression text types in PDMX. Tempo and dynamic markings are the most common, as every song requires at least one of each, while pedal and text spanners are the least frequent.

Figure 3: Expression Text Durations . Boxplots of durations (in beats) for different categories of expression text. Key signatures, rehearsal marks, and time signatures tend to span entire sections of songs, and thus have some of the longest durations. Pedal and text spanners, meanwhile, usually only apply to short sequences of notes and so are often quite brief. Tempo markings exhibit a long tail of extended durations, likely because many songs contain frequent tempo changes resulting in short local spans, while a few songs may have an initial tempo that persists across the entire piece.

Symbolic Music Generation

Symbolic music generation creates and edits music as structured notes, rhythms, chords, and scores rather than directly as audio.

See all 58 papers

TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description
Oct 11, 2026Mingyang Yao, Jingyue Huang, Ke Chen +2Text-to-Music Generation Symbolic Music Generation
University of California, San Diego · University of Rochester
Text2Score: Generating Sheet Music From Textual Prompts
May 13, 2026Keshav Bhandari, Sungkyun Chang, Abhinaba Roy +4Text-to-Music Generation Music Generation Evaluation
Queen Mary University of London, UK · EmotionWave, Korea · Singapore University of Technology and Design, Singapore +1
Getting Motif-ated: Controllable AI Compositions from Injected Motif Prompts
Sep 26, 2026Chao Peter Yang, Cynthia Rudin, Yue Jiang +2Controllable Music Generation Symbolic Music Generation
Yale University New Haven, CT · Duke University Durham, NC · Chinese University of Hong Kong, Shenzhen Shenzhen, Guangdong

Oct 11, 2026J/K move · Enter open · S save

TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description

Mingyang Yao, Jingyue Huang, Ke Chen, Taylor Berg-Kirkpatrick+1

University of California, San Diego · University of Rochester

Text-to-Music Generation Symbolic Music Generation Controllable Music Generation

Open paper AskPDF arXiv

Text-conditioned symbolic music generation has attracted growing interest with the development of generative models. Yet two questions remain underexplored: what aspects of music the conditioning text should describe, and how the corresponding data should be collected and trained to improve controllability and text-music alignment. We address these challenges with TASTE, a taxonomy-aware framework for symbolic music generation with a main focus on piano pieces. First, we introduce a taxonomy of musical elements that specifies how a piece can be captured through textual description in a musically grounded way. The taxonomy comprises a set of metadata tags (genre, emotion, section plan, etc.) together with a free-form description of music imagery. This structure decomposes textual input to match the generation goal more coherently and precisely. Second, we build a data pipeline for collecting, annotating, and retrieving public symbolic piano resources, yielding a corpus of 432K MIDI files labeled at multiple levels of granularity. Finally, we design a staged training procedure that leverages this multi-granularity data across two model families: an autoregressive model and a non-autoregressive diffusion model. Both objective and subjective evaluations show that our models achieve superior musical-element control, text-description alignment, and overall music quality. We will release the full pipeline, all annotated datasets, and the model checkpoints.

Abstract

Symbolic Music Generation

Explore similar work

TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description

Text2Score: Generating Sheet Music From Textual Prompts

Getting Motif-ated: Controllable AI Compositions from Injected Motif Prompts