Transcribing expressive piano performances into scores is tricky because note timings must be interpreted within an underlying metrical structure, and beats often occur without coincident note onsets. Existing systems typically treat note-level and frame-level information separately, which can lead to errors in score timing and notated duration.

DuET, introduced by Wei Zeng and Ye Wang, is a dual-encoder transformer that learns both note-level performance MIDI-to-score transcription and frame-level beat and downbeat tracking. Bidirectional cross-attention connects the note and frame encoders, and an autoregressive decoder alternates between their representations to predict structured score attributes. The authors also propose TimeRoPE and Time-Pitch Rotary Positional Encoding (TPRoPE), which establish a shared performance-time reference and explicitly encode relative temporal and pitch relationships.

The model was evaluated on the ASAP classical piano dataset and a newly curated Gangqinpu dataset spanning multiple genres with a focus on popular piano music. Compared with the strongest baseline, DuET reduces MUSTER mean error from 6.56% to 3.64% on Gangqinpu and from 11.30% to 11.22% on ASAP. Ablations confirm the benefits of frame-level context, time-based positional encoding, and decoder-history masking. Code, weights, and demos are available online.