D3PIA: A Discrete Denoising Diffusion Model for Piano Accompaniment Generation From Lead sheet

TL;DR

The paper tackles piano-accompaniment generation from lead sheets by formulating the task as discrete diffusion on a 4-state piano-roll, guided by a local lead-sheet conditioning mechanism. It introduces D3PIA, which combines a Neighborhood Attention–based lead-sheet encoder with a NA-driven denoising decoder, operating at 100 diffusion steps and employing absorbing-state sampling for refinement. Evaluations on POP909 show that D3PIA achieves superior harmonic coherence and rhythmic consistency compared with continuous-diffusion and Transformer baselines, while offering faster inference and a smaller model footprint. The work highlights the value of discrete-state modeling and locality-aware conditioning for controllable, high-quality accompaniment generation, and discusses future extensions to velocity modeling and richer style controls.

Abstract

Generating piano accompaniments in the symbolic music domain is a challenging task that requires producing a complete piece of piano music from given melody and chord constraints, such as those provided by a lead sheet. In this paper, we propose a discrete diffusion-based piano accompaniment generation model, D3PIA, leveraging local alignment between lead sheet and accompaniment in piano-roll representation. D3PIA incorporates Neighborhood Attention (NA) to both encode the lead sheet and condition it for predicting note states in the piano accompaniment. This design enhances local contextual modeling by efficiently attending to nearby melody and chord conditions. We evaluate our model using the POP909 dataset, a widely used benchmark for piano accompaniment generation. Objective evaluation results demonstrate that D3PIA preserves chord conditions more faithfully compared to continuous diffusion-based and Transformer-based baselines. Furthermore, a subjective listening test indicates that D3PIA generates more musically coherent accompaniments than the comparison models.

Figures (2)

• Figure 1: D3PIA model structure. Lead sheet information is attended to in the encoder, and the encoder output (shown in dark green) is concatenated (denoted as C) with the noisy piano roll and further attended to in the denoising decoder via NA. N denotes the number of NA layers. In the NA 2D Self module, T, D, and P refer to timestep, embedding dimension, and pitch, respectively, while Q, K, and V represent the query, key, and value used for attention computation. $\tau$ indicates the diffusion timestep.
• Figure 2: Subjective evaluation result with 95 % confidence interval.