Research map / Image generation / Autoregressive and latent generation
Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a two-stage music arrangement system that models a style prior over disentangled piano-texture and orchestral-function codes, using a layer-interleaved multi-stream Transformer to turn lead sheets into structured multi-track accompaniments.
Problem
Arranging structured, multi-track band accompaniments from a lead sheet while keeping track cohesion, long-term structural coherence, and computational efficiency.
Approach
Model a prior p(style factors | content factors) over disentangled piano-reduction content and VQ-coded orchestral-function style, with a multi-stream Transformer that interleaves time-wise and track-wise layers, so track count and instrumentation are flexible and cost scales with max(N,T) rather than NT.
Key results
- On lead-sheet-to-multi-track arrangement, the model reaches chord accuracy 0.567 and structure 1.520, versus 0.470 and 1.086 for the best and weakest baselines in those columns, with latency 0.461 s/bar, close to GETMusic's 0.450.
- Latency is roughly 10x faster than AMT (6.320 s/bar) while achieving higher chord accuracy and structure awareness.
- In a 23-participant survey, the method was rated highest on all five criteria, with significant differences across criteria (p < 0.05).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed two-stage style-prior arrangement system instead of existing baselines | Lead sheet to multi-track arrangement · Nottingham and WikiMT | chord accuracy: 0.470±0.013 -> 0.567±0.014 | Helped |
| use the proposed two-stage style-prior arrangement system instead of existing baselines | Lead sheet to multi-track arrangement · Nottingham and WikiMT | DOA: 0.270±0.007 -> 0.300±0.004 | Helped |
| use the proposed two-stage style-prior arrangement system instead of existing baselines | Lead sheet to multi-track arrangement · Nottingham and WikiMT | structure awareness: 1.243±0.017 -> 1.520±0.030 | Helped |
| use the proposed system instead of note-level autoregressive AMT | Lead sheet to multi-track arrangement · Nottingham and WikiMT | inference latency (second/bar): 6.320±0.212 -> 0.461±0.005 | Helped |
| use the proposed system instead of GETMusic | Lead sheet to multi-track arrangement · Nottingham and WikiMT | inference latency (second/bar): 0.450±0.002 -> 0.461±0.005 | Hurt |
| replace interleaved layer architecture with Parallel stream embedding (ablation) | Piano-to-multi-track prior · POP909 test set | faithfulness (latent): 0.215±0.005 -> 0.153±0.003 | Hurt |
| replace interleaved layer architecture with Delay code interleaving (ablation) | Piano-to-multi-track prior · POP909 test set | faithfulness (latent): 0.215±0.005 -> 0.133±0.003 | Hurt |
| replace Stage 1 Piano Texture Prior with Whole-Song-Gen (ablation) | Two-stage arrangement system · Nottingham and WikiMT | chord accuracy: 0.567±0.014 -> 0.509±0.015 | Hurt |
| replace Stage 1 Piano Texture Prior with Whole-Song-Gen (ablation) | Two-stage arrangement system · Nottingham and WikiMT | structure awareness: 1.520±0.030 -> 1.121±0.006 | Hurt |
| replace Stage 1 Piano Texture Prior with Whole-Song-Gen (ablation) | Two-stage arrangement system · Nottingham and WikiMT | DOA: 0.300±0.004 -> 0.277±0.006 | Hurt |
| replace Piano Texture Prior with Whole-Song-Gen (ablation) | Lead sheet to piano module · Lead sheet to piano | chord accuracy: 0.540±0.016 -> 0.430±0.020 | Hurt |
| replace Piano Texture Prior with Whole-Song-Gen (ablation) | Lead sheet to piano module · Lead sheet to piano | structure awareness: 1.983±0.147 -> 1.153±0.180 | Hurt |
Surprising findings
- The paper reports the stated DOA ground truth (0.333) but the proposed model's DOA (0.300) is below it, so creativity is not matched to human-written data.
Limitations
- Supports only tonal tracks in quadruple meter; excludes triple meter, triplets, and drums.
- Models composition level only; no MIDI velocity, timing, or control messages.
- Trained and evaluated mainly on contemporary Western music, a cultural bias.
Experimental setup
Datasets: Slakh2100 Lakh MIDI Dataset Nottingham WikiMT POP909
Benchmarks: Nottingham WikiMT POP909
Metrics: chord accuracy degree of arrangement structure awareness (ILS) inference latency faithfulness NLL
Baselines: PopMAG Anticipatory Music Transformer GETMusic Whole-Song-Gen Parallel multi-stream Delay multi-stream
Builds on: VQ-VAE PianoTree VAE Anticipatory Music Transformer Transformer-XL AccoMontage
Closest papers by meaning
- Unifying Symbolic Music Arrangement: Track-Aware Reconstruction and Structured Tokenization (NeurIPS 2025 · similarity 0.90)
- JEN-1 Composer: A Unified Framework for High-Fidelity Multi-Track Music Generation (AAAI 2025 · similarity 0.90)
- MIDI-GPT: A Controllable Generative Model for Computer-Assisted Multitrack Music Composition (AAAI 2025 · similarity 0.88)
- Compose with Me: Collaborative Music Inpainter for Symbolic Music Infilling (AAAI 2025 · similarity 0.87)
- MuPT: A Generative Symbolic Music Pretrained Transformer (ICLR 2025 · similarity 0.86)
- Whole-Song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.