Read-only mirror of the MIREX Wiki, snapshot of 19 November 2026. Open this page on the original wiki to edit it or to see newer changes.

2026:Audio-to-Score Transcription Results

From MIREX Wiki

Submissions[edit]

Team PDF A2S Track Training size Number of parameters Training GPU hours Inference real-time factor
Lu et al. PDF / Poster Staves-Informed 24,333 (Quartets) 47.1M (46.5M trainable) 22.92 (H800) 0.690
Cummins et al. PDF Blind Quartets
Hidalgo-Centeno et al. PDF Blind 24,351 (Quartets) 322M 185 (RTX 5090 32GB) 0.408

The code and data used to evaluate the submissions is available on the Public evaluation repository.

QuartetsMIREX 2026 test set[edit]

The "Undisclosed Evaluation Set" mentioned in the challenge consists of 8 string quartets score taken from [kern.ccarh.org](https://kern.ccarh.org) and processed like the original Quartets dataset to produce 373 kern excerpts. These quartets are from composers not represented in the original dataset: Schubert, Brahms, and Mendelssohn.

Staves-Informed A2S[edit]

Quartets test set[edit]

Submission WER CER LER MV2H
Lu et al. 0.2160 0.2385 0.5983 0.8796

QuartetsMIREX 2026[edit]

Submission WER CER LER MV2H
Lu et al. 0.3177 0.3481 0.7456 0.8179

Blind A2S[edit]

Quartets test set[edit]

Submission WER CER LER MV2H
Cummins et al. 0.0346 0.0347 0.0980 0.9478
Hidalgo-Centeno et al. 0.0286 0.0276 0.0792 0.9492

QuartetsMIREX 2026[edit]

Submission WER CER LER MV2H
Cummins et al. 0.3202 0.3150 0.7679 0.5981
Hidalgo-Centeno et al. 0.3385 0.3258 0.7460 0.6053