Drum transcription usually depends on task-specific supervision, and Analysis-by-Synthesis (AbS) models offer a flexible alternative by reconstructing audio from one-shot events. However, AbS systems often struggle with unusual timbres such as electronic drum kits, need high-quality paired audio-transcription data, and fail to capture intra-class variation like multiple kick sounds in one arrangement.

D-Composer, proposed by researchers at ISMIR 2026, recasts the problem as language modeling. An autoregressive model consumes compact acoustic token representations of a drum recording and outputs labeled sound events with onset times. Trained on abundant synthetic data, it learns to decompose audio without relying on curated real-world pairs.

In evaluations, D-Composer performed comparably to or better than a state-of-the-art AbS system and task-specific baselines for drum transcription and one-shot extraction, while generalizing across a broader range of drum timbres, including electronic kits. The authors argue this makes language modeling a simple, unified recipe for transcription and sound event decomposition.