Structure-Aware Piano Accompaniment via Style Planning and Dataset-Aligned Pattern Retrieval

Wanyu Zang; Yang Yu; Meng Yu

cs.SDFeb 16, 2026

Structure-Aware Piano Accompaniment via Style Planning and Dataset-Aligned Pattern Retrieval

Authors: Wanyu Zang, Yang Yu, Meng Yu

Organizations: Computer Science, Lewis University · Nexus Art Foundation · Computer Science, Governors State University

Abstract

We introduce a structure-aware approach for symbolic piano accompaniment that decouples high-level planning from note-level realization. A lightweight transformer predicts an interpretable, per-measure style plan conditioned on section/phrase structure and functional harmony, and a retriever then selects and reharmonizes human-performed piano patterns from a corpus. We formulate retrieval as pattern matching under an explicit energy with terms for harmonic feasibility, structural-role compatibility, voice-leading continuity, style preferences, and repetition control. Given a structured lead sheet and optional keyword prompts, the system generates piano-accompaniment MIDI. In our experiments, transformer style-planner-guided retrieval produces diverse long-form accompaniments with strong style realization. We further analyze planner ablations and quantify inter-style isolation. Experimental results demonstrate the effectiveness of our inference-time approach for piano accompaniment generation.

Figures & tables

Approach	Realization	Retr.	Structure	Style	Planning
Autoregressive note-level [ 1 ] and [ 2 ]	Note or event gen.	N	Implicit	Priming/cond.	None
Beat-based modeling (REMI) [ 10 ]	Note/event gen.	N	Implicit	Representation	None
Anticipatory infilling [ 8 ]	Note or event gen.	N	Control events	Control tokens	None
Diffusion inpainting/controls [ 4 ]	Note or event gen.	N	Local/inpaint	Conditioning	None
Whole-song hierarchical diffusion [ 3 ]	Note or event gen.	N	Explicit hierarchy	Conditioning	Hierarchical
Style-prior arrangement [ 11 ]	Hybrid retr.+gen.	Y	Phrase and section	Style priors	Transformer prior

TABLE I: Representative comparisons along key axes for symbolic accompaniment and arrangement. “Structure” indicates whether explicit section and phrase cues are used rather than implicit or learned structure. “Planning” indicates whether a distinct song-level plan is produced before note-level realization.

Fig. 1: Overview of our style-planning and retrieval-based accompaniment pipeline. Offline preparation derives discrete measure-level style labels (dynamics, articulation, rhythm, tension, texture, register) and builds a measure-level index from POP909. Online generation predicts per-measure style slots with a Transformer style planner, then retrieves and adapts dataset patterns to the target harmony with voice-leading-aware selection. Solid arrows indicate runtime flow; dashed arrows indicate offline assets.

Song	Ent norm	Top-3	Dom.	Miss	Pool
dngl	1.000	0.399	0.166	5	12
gsn	1.000	0.218	0.073	6	29
lbkch	0.996	0.274	0.110	4	13
rsj	1.000	0.297	0.132	2	31
ts	0.750	0.352	0.284	3	31
Mean $\pm$ Std	$0.949\pm 0.100$	$0.308\pm 0.063$	$0.153\pm 0.072$	$4.0\pm 1.4$	$23.2\pm 8.8$

TABLE II: Section-wise diversity and retrieval health. Ent norm is normalized entropy. Top-3 is top-3 coverage. Dom. is dominant-cluster ratio. Pool is the median strict candidate pool size for style+QUAL. Miss is the count of measures with an empty strict pool, so relaxation is used.

Metric	Definition, brief	Fixed style	Style planner	$\Delta$ , 95% CI
Transition distance $\uparrow$	Mean adjacent-measure style change.	$0.000\pm 0.000$	$0.309\pm 0.019$	$+0.309$ $[0.294,0.326]$
Boundary contrast $\uparrow$	Boundary dist. $-$ within-section dist.	$0.000\pm 0.000$	$0.223\pm 0.036$	$+0.223$ $[0.195,0.259]$
Axis entropy $\uparrow$	Mean per-axis entropy within sections.	$0.000\pm 0.000$	$0.720\pm 0.044$	$+0.720$ $[0.681,0.761]$
Vector entropy $\uparrow$	Entropy of full style vectors within sections.	$0.000\pm 0.000$	$3.116\pm 0.326$	$+3.116$ $[2.817,3.395]$
Section mode ratio $\downarrow$	Share of dominant style vector per section.	$1.000\pm 0.000$	$0.408\pm 0.057$	$-0.592$ $[-0.646,-0.547]$
Axis consistency $\downarrow$	Mean per-axis mode share within sections.	$1.000\pm 0.000$	$0.758\pm 0.024$	$-0.242$ $[-0.264,-0.221]$

TABLE III: Ablation results. We compare a fixed-style baseline without a planner against a learned style planner using style-space metrics derived from target style labels and section structure. Numbers are mean

\pm

std across songs. We also report 95% bootstrap CIs with n=5 songs for

\Delta=\text{planner}-\text{fixed}

. All CIs exclude zero, indicating statistically reliable planner effects in style space.

Song	Measures	Active pairwise	Active union	Pitch pairwise	Pitch union
dngl	79	0.0215	672	0.0209	770
gsn	173	0.0214	969	0.0215	1191
lbkch	99	0.0223	826	0.0220	963
rsj	127	0.0137	960	0.0137	1156
ts	121	0.0199	949	0.0203	1106
Mean $\pm$ Std	$119.8\pm 31.6$	$0.0198\pm 0.0031$	$875.2\pm 114.1$	$0.0197\pm 0.0030$	$1037.2\pm 154.5$

TABLE IV: Inter-style isolation. Pairwise is the mean pairwise match. Union is the number of distinct signatures within the song. The last row reports mean

\pm

std across songs.

Style	Starts/Meas	Vel med	Staccato	RegSpan
arp	$9.51\pm 0.33$	$73.0\pm 4.4$	$0.093\pm 0.026$	$25.0\pm 1.3$
block	$9.19\pm 0.39$	$73.4\pm 5.2$	$0.111\pm 0.045$	$26.6\pm 1.7$
busy	$15.36\pm 0.76$	$75.4\pm 4.3$	$0.287\pm 0.054$	$26.2\pm 1.6$
dense	$15.19\pm 1.18$	$74.0\pm 4.7$	$0.228\pm 0.055$	$25.4\pm 1.5$
fast	$16.74\pm 0.86$	$73.8\pm 4.4$	$0.230\pm 0.050$	$26.2\pm 1.6$
gentle	$11.18\pm 0.86$	$74.4\pm 5.1$	$0.057\pm 0.014$	$25.6\pm 1.5$

TABLE V: Realized continuous features.

Panel of Figure 7 — Fig. 2: Training dynamics of the Transformer style planner (20 epochs). The model converges quickly, with validation loss and accuracy peaking early.

Music Accompaniment Generation

See all 12 papers

Learning Music Style for Piano Arrangement Through Cross-Modal Bootstrapping
Aug 4, 2026Jingwei Zhao, Gus Xia, Ziyu Wang +1Controllable Music Generation Cross-Modal Alignment
1Songscription · School of Computing, National University of Singapore · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) +1
Learning Jazz Pianist Style with Cross-Attention Conditioning
Oct 2, 2026Drew Edwards, Akira Maezawa, Simon DixonMusic Generation Symbolic Music Generation
Centre for Digital Music, Queen Mary University of London · Yamaha Corporation Hamamatsu, Japan
HAFM: Hierarchical Autoregressive Foundation Model for Music Accompaniment Generation
Apr 10, 2026Jian Zhu, Jianwei Cui, Yunlong Xue +4Controllable Music Generation Music Generation
Zhejiang Lab · University of Science and Technology of China · Zhejiang International Studies University

Aug 4, 2026J/K move · Enter open · S save

Learning Music Style for Piano Arrangement Through Cross-Modal Bootstrapping

Jingwei Zhao, Gus Xia, Ziyu Wang, Ye Wang

1Songscription · School of Computing, National University of Singapore · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)+1

Controllable Music Generation Cross-Modal Alignment Symbolic Music Generation Music Generation

Open paper AskPDF arXiv

What is music style? Though often described using text labels such as "swing," "classical," or "emotional," the real style remains implicit and hidden in concrete music examples. In this paper, we introduce a cross-modal framework that learns implicit music styles from raw audio and applies them to symbolic music generation. Inspired by BLIP-2, our model leverages a Querying Transformer (Q-Former) to extract style representations from a large, pre-trained audio language model (LM), and further applies them to condition a symbolic LM for generating piano arrangements. We adopt a two-stage training strategy: contrastive learning to align auditory style with symbolic expression, followed by generative modeling for music arrangement. Our model generates piano performances jointly conditioned on a lead sheet (content) and a reference audio example (style), enabling controllable and stylistically faithful arrangement. Experiments demonstrate the effectiveness of our approach in piano cover generation, style transfer, and audio-to-MIDI retrieval, achieving substantial improvements in style-aware alignment and music quality.

Abstract

Music Accompaniment Generation

Explore similar work

Learning Music Style for Piano Arrangement Through Cross-Modal Bootstrapping

Learning Jazz Pianist Style with Cross-Attention Conditioning

HAFM: Hierarchical Autoregressive Foundation Model for Music Accompaniment Generation