This paper presents the results of the 2025 Automatic Music Transcription (AMT) Challenge, an online competition to benchmark progress in multi-instrument transcription. Eight teams submitted valid solutions; two outperformed the baseline MT3 model. The results highlight both advances in transcription accuracy and the remaining difficulties in handling polyphony and timbre variation. We conclude with directions for future challenges: broader genre coverage and stronger emphasis on instrument detection.
Figures & tables
Figure 1: Excerpts of the sample music available to the participants.
Figure 2: End-to-end data flow of the transcription challenge.
Figure 3: Evaluation Method. Top: Reference. Bottom: Estimated notes.
Rank
Model Name
F1 Score
Precision
Recall
Overlap
Runtime
1
MIROS
0.5998
0.6558
0.5724
0.7391
22.05
2
YourMT3-YPTF-MoE-M
0.5938
0.6010
0.5888
0.7305
12.60
3
YourMT3-YPTF-S
0.5581
0.5565
0.5615
0.7326
15.40
4
YourMT3-P
0.3947
0.3966
0.3985
0.7263
14.99
5
MT3 [ 9 ] (baseline)
0.3932
0.3811
0.4115
0.7180
20.19
6
YourMT3-YPTF-SP-V
0.3305
0.3280
0.3358
0.7147
14.50
Table 1: Results from all evaluated systems. Multiple variants and public baselines are included; MT3 serves as the reference baseline. Runtime is measured by ms.
Model
Instrument Count
F-measure
Precision
Recall
MT3 (Baseline)
Count = 1 (n=6)
0.5483 ± 0.3898
0.5548 ± 0.3974
0.5422 ± 0.3829
MT3 (Baseline)
Count = 2 (n=24)
0.3792 ± 0.1829
0.3661 ± 0.1808
0.3966 ± 0.1894
MT3 (Baseline)
Count = 3 (n=46)
0.2559 ± 0.1388
0.2475 ± 0.1409
0.2681 ± 0.1383
MIROS
Count = 1 (n=6)
0.7193 ± 0.2103
0.9067 ± 0.2286
0.6233 ± 0.2295
MIROS
Count = 2 (n=24)
0.5004 ± 0.2403
0.5460 ± 0.2296
0.4749 ± 0.2530
MIROS
Count = 3 (n=46)
0.4367 ± 0.2012
0.4643 ± 0.2044
0.4255 ± 0.2101
Table 2: Performance comparison across exact instrument counts (1, 2, and 3 instruments) for the top two models alongside the MT3 baseline. Values are reported as mean ± standard deviation. Welch’s t-test results are shown above in the text. The variable n denotes the number of evaluated pieces belonging to each instrument count category.
Existing methods for automatic music transcription are often limited to single-instrument recordings or fail on complex, real music mixes. Although previous work utilizes synthetic training data, the resulting models generalize poorly, leading to largely unusable transcription output in realistic, multi-instrument settings. In this work, we analyze the effectiveness of synthetic data for pre-training while combining it with fine-tuning on real music audio and post-training using reinforcement learning. We further introduce conditioning on instrument presence to customize transcriptions. Finally, we release MuScriptor, an open-weight multi-instrument music transcription model that works on real-world music recordings from across a diverse range of musical genres.
Simon Rouard, Michael Krause, Axel Roebel +2
Kyutai · Mirelo AI · UMR STMS, IRCAM-CNRS Sorbonne Univ.
Automatic Music Transcription (AMT) for guitar remains limited by three challenges: existing systems often fail to capture expressive techniques such as slides, bends, and percussive hits; they often assign notes to incorrect string-fret combinations; and they are typically trained on clean recordings, limiting their generalization to noisy real-world audio. To address these challenges, we propose TART, a modular four-stage audio-to-tablature pipeline consisting of (1) an audio-to-MIDI transcription model, (2) an expressive technique classifier, (3) an audio-conditioned T5 encoder-decoder for string-fret assignment, and (4) an automated tablature generator. We evaluate TART in a zero-shot setting on GuitarSet, EGDB, and two augmented benchmarks, Noisy GuitarSet and Noisy EGDB. Averaged across these four benchmarks, TART achieves 81.35% audio-to-MIDI F50 (+6.67 points over the best prior baseline), 71.8% string-fret Tab F1 (+8.5 points over the best prior baseline), and 54.08% end-to-end Tab F1. To our knowledge, TART is the first framework to generate guitar tablature with both fingering and expressive technique annotations directly from guitar audio.
Competitive music transcription models require large amounts of paired audio-score data, which is scarce due to collection costs, alignment difficulty, and copyright restrictions. Meanwhile, vast quantities of unpaired audio recordings and symbolic scores are freely available but have gone unused. We adopt a cycle-consistent translation framework in which a small amount of paired data acts as a minimal anchor, unlocking the full potential of the unpaired pool. We find that: unpaired data yields surprisingly large gains, especially under limited supervision; unpaired audio contributes more than unpaired scores; incorporating unlabeled audio from a new instrument during training improves transcription for that instrument without any paired supervision. Together, these results suggest that scaling unpaired data offers a practical path toward high-quality transcription for instruments where labeled data remains scarce.