D3PIA: A Discrete Denoising Diffusion Model for Piano Accompaniment Generation from Lead Sheet

Submitted to ICASSP 2026

In this page, we demonstrate our proposed discrete diffusion-based piano accompaniment generation model, D3PIA, leveraging the locally aligned structure of musical accompaniments with the lead sheet in the piano roll representation. As mentioned in the Results & Discussions Section, we also present the objective and subjective scores of samples.

🎵 Select a Sample

🤖 Choose a Model

Loading...

D3PIA

Trained with 8 bars. Temperature = 1.5

GT

Test split of POP909 Dataset including melody, bridge, and arrangement track

Polyffusion

Trained with 8 bars. Generated with inpainting "below" option. chord CFG cond_scale = 5.0

C&E-E

Re-implemented (GPT-2) embellish part. Temp = 1.1, p = 0.95

WSG-4th

Stage 4 model of WholeSongGen.

FGG

Trained with 8 bars (4 bars in original code).

Leadsheet

Melody and chord. Segments selected when melody exists in ≥4 bars.

Scores

Objective Metrics Subjective Metrics
Model Chord Accuracy Chord Similarity Harmony Correctness