Scoring Time Intervals Using Non-Hierarchical Transformer for Automatic Piano Transcription
ISMIR - 2024
The TransKun V2 paper reports a transformer interval-scoring architecture and a higher F1 result on Maestro subtasks.
Loading...
Turn piano recordings into editable MIDI for notation, practice and arranging.
Input
Clear recordings of solo piano or a single keyboard instrument.
Supported formats
Output
Piano MIDI generated by TransKun V2 Aug, for comparison with V2 results.
Use case
Compare its transcription with TransKun V2 on the same recording to decide which version better matches the performance.
Choose a model for your material and purpose, and review the examples before submitting.
Cost, time, formats and parameters depend on the selected model.
Include the tool name, steps and the problem in your feedback. Screenshots can help us locate it.
TransKun V2 Aug is a separate local piano transcription route using the official data-augmentation checkpoint. Submit a piano file to generate MIDI without replacing the default V2 tool.
TransKun V2 Aug is the separate local data-augmentation checkpoint route using checkpointTransformerAug.zip.
Uses the official Standard checkpoint, whose note offsets follow pedal-extended acoustic duration. It is kept separate from the key-release-oriented NoExt and the V2 Aug checkpoint.
The local default is the upstream transkun package's TransKun V2 No Pedal Ext checkpoint (2.0.pt), in the same architecture family and fine-tuned and extended through the V2 Aug route.
Aria-AMT uses the EleutherAI seq-to-seq piano route with piano-medium-double-1.0.safetensors.
Transcribe piano audio into discrete note events with the released D3RM diffusion-refinement checkpoint, then export editable MIDI and note-event data.
ByteDance Piano CRNN note+pedal uses note_F1=0.9677_pedal_F1=0.9186.pth for solo-piano MIDI with pedal controller events.
SFT-CRNN detects piano note onsets and offsets, with fixed velocity 100 and no pedal prediction.
ISMIR - 2024
The TransKun V2 paper reports a transformer interval-scoring architecture and a higher F1 result on Maestro subtasks.
NeurIPS - 2021
The 2021 Semi-CRF paper frames piano transcription as interval/event prediction instead of stitching frame-level detections.
Official repository / preliminary weights - current
The Aria-AMT repository documents the seq-to-seq piano transcription implementation and preliminary weights; the paper reports MAESTRO/MAPS Table 1 comparisons.
arXiv - 2025
Transcribe piano audio into discrete note events with the released D3RM diffusion-refinement checkpoint, then export editable MIDI and note-event data.
arXiv - 2020
The ByteDance paper reports 96.72% note onset F1 and 91.86% pedal onset F1 on MAESTRO, with released code and checkpoints.
ICASSP - 2026
SFT-CRNN detects piano note onsets and offsets, with fixed velocity 100 and no pedal prediction.
Includes public test results for models used here and elsewhere. The complete source table is shown by default; conditions and metrics can also be selected for comparison. Scores from different test conditions are not directly comparable.
Only records with verified public inference implementations and matching weights are shown; algorithms requiring no weights are also eligible. Closed services, unreleased weights and unmatched experiment versions are excluded. This is not an all-model leaderboard.
Public weights do not establish a fully open-source license. Expand the data for declared weight licenses and access conditions. Undeclared licenses are marked separately.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
wlazbzfll · MIREX 2024 summary · teamWLY · MIREX 2024 summary · hFT-Transformer · MIREX 2024 summary
Verified public/reproducible MIREX rows compare average note-onset F1 across three datasets and holistic note F1 across Maestro and SMD.
| Metric rank | Model | Test setting | Summary onset F1 (%)↑ | Holistic note F1 (%)↑ |
|---|---|---|---|---|
| 1 | TransKun V2 AugUsed on this sitePublic weights · Inference code | MIREX 2024 summary | 96.48 | 90.81 |
| 2 | TransKun V2Public weights · Inference code | MIREX 2024 summary | 94.9 | 87.64 |
Markers identify models or components used here. Scores and settings come from the cited tests, not a new evaluation of the complete site pipeline.
Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.
Deployable entries 9 / source rows 9
Deployable entries 2 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
wlazbzfll · MIREX 2024 summary · teamWLY · MIREX 2024 summary · hFT-Transformer · MIREX 2024 summary
Deployable entries 2 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
wlazbzfll · MIREX 2024 published row · teamWLY · MIREX 2024 published row · hFT-Transformer · MIREX 2024 published row
Deployable entries 2 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
wlazbzfll · MIREX 2024 published row · teamWLY · MIREX 2024 published row · hFT-Transformer · MIREX 2024 published row
Deployable entries 2 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
wlazbzfll · MIREX 2024 published row · teamWLY · MIREX 2024 published row · hFT-Transformer · MIREX 2024 published row
Deployable entries 0 / source rows 10
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Aria-AMT + bootstrapping · MAESTRO · paper Table 1 · Aria-AMT · MAESTRO · paper Table 1 · hFT-Transformer · MAESTRO · paper Table 1 · Google · MAESTRO · paper Table 1 · Kong · MAESTRO · paper Table 1 · Aria-AMT + bootstrapping · MAPS · paper Table 1 · Aria-AMT · MAPS · paper Table 1 · hFT-Transformer · MAPS · paper Table 1 · Kong · MAPS · paper Table 1 · Google · MAPS · paper Table 1
Deployable entries 2 / source rows 7
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
HPPNet-sp · HFT-transformer · Yan et al. · PAR · DiffRoll (k=9)
Deployable entries 1 / source rows 6
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Onsets & frames · Onsets & frames [reproduced] · Adversarial onsets & frames · Regress cond on. · Regress cond on. & off.
Deployable entries 1 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Liang · Onsets & frames [our-implemented] · Onsets & frames (misaligned labels) · proposed (misaligned labels)
Deployable entries 1 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Onsets & Frames · HPPNet-sp · hFT-Transformer · Transkun
Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Verified public/reproducible MIREX rows compare average note-onset F1 across three datasets and holistic note F1 across Maestro and SMD.
| Metric rank | Model | Test setting | Summary onset F1 (%)↑ | Holistic note F1 (%)↑ |
|---|---|---|---|---|
| 1 | TransKun V2 AugPublic weights · Inference code | MIREX 2024 summary | 96.48 | 90.81 |
| 2 | TransKun V2Public weights · Inference code | MIREX 2024 summary | 94.9 | 87.64 |
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Official MIREX MAESTRO results, including all published note and pedal P / R / F1 (%); unreported metrics remain absent.
| Metric rank | Model | Test setting | Activation P (%)↑ | Activation R (%)↑ | Activation F1 (%)↑ | Note onset P (%)↑ | Note onset R (%)↑ | Note onset F1 (%)↑ | Note onset + offset P (%)↑ | Note onset + offset R (%)↑ | Note onset + offset F1 (%)↑ | Note onset + offset + velocity P (%)↑ | Note onset + offset + velocity R (%)↑ | Note onset + offset + velocity F1 (%)↑ | Pedal activation P (%)↑ | Pedal activation R (%)↑ | Pedal activation F1 (%)↑ | Pedal onset P (%)↑ | Pedal onset R (%)↑ | Pedal onset F1 (%)↑ | Pedal onset+offset P (%)↑ | Pedal onset+offset R (%)↑ | Pedal onset+offset F1 (%)↑ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | TransKun V2Public weights · Inference code | MIREX 2024 published row | 95.76 | 94.89 | 95.3 | 99.56 | 97.14 | 98.32 | 94.65 | 92.38 | 93.49 | 94.11 | 91.86 | 92.96 | 96.71 | 94.53 | 95.41 | 89.09 | 84.21 | 86.42 | 86.32 | 81.65 | 83.77 |
| 2 | TransKun V2 AugPublic weights · Inference code | MIREX 2024 published row | 94.95 | 95.22 | 95.05 | 99.71 | 97.15 | 98.4 | 94.37 | 91.97 | 93.14 | 93.86 | 91.49 | 92.64 | 95.46 | 94.16 | 94.54 | 88.83 | 81.16 | 84.53 | 84.97 | 77.9 | 81.02 |
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Official MIREX MAPS results, including all published note and pedal P / R / F1 (%); unreported metrics remain absent. MAPS uses ad hoc alignment; interpret offset metrics in that alignment context.
| Metric rank | Model | Test setting | Activation P (%)↑ | Activation R (%)↑ | Activation F1 (%)↑ | Note onset P (%)↑ | Note onset R (%)↑ | Note onset F1 (%)↑ | Note onset + offset P (%)↑ | Note onset + offset R (%)↑ | Note onset + offset F1 (%)↑ | Note onset + offset + velocity P (%)↑ | Note onset + offset + velocity R (%)↑ | Note onset + offset + velocity F1 (%)↑ | Pedal activation P (%)↑ | Pedal activation R (%)↑ | Pedal activation F1 (%)↑ | Pedal onset P (%)↑ | Pedal onset R (%)↑ | Pedal onset F1 (%)↑ | Pedal onset+offset P (%)↑ | Pedal onset+offset R (%)↑ | Pedal onset+offset F1 (%)↑ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | TransKun V2 AugPublic weights · Inference code | MIREX 2024 published row | 94.46 | 83.34 | 88.43 | 93.96 | 90.56 | 92.19 | 71.05 | 68.54 | 69.75 | 55.96 | 54.01 | 54.95 | 88.93 | 83.89 | 85.83 | 73.13 | 75.29 | 73.43 | 54.99 | 56.5 | 55.32 |
| 2 | TransKun V2Public weights · Inference code | MIREX 2024 published row | 88.7 | 82.52 | 85.35 | 86.71 | 90.48 | 88.49 | 63.25 | 66.13 | 64.61 | 43.51 | 45.51 | 44.46 | 84.98 | 85.47 | 84.49 | 65.21 | 71.82 | 67.32 | 49.03 | 54.27 | 50.88 |
The table includes only models with public weights or configurations that can run locally.
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Official MIREX SMD results, including all published note and pedal P / R / F1 (%); unreported metrics remain absent.
| Metric rank | Model | Test setting | Activation P (%)↑ | Activation R (%)↑ | Activation F1 (%)↑ | Note onset P (%)↑ | Note onset R (%)↑ | Note onset F1 (%)↑ | Note onset + offset P (%)↑ | Note onset + offset R (%)↑ | Note onset + offset F1 (%)↑ | Note onset + offset + velocity P (%)↑ | Note onset + offset + velocity R (%)↑ | Note onset + offset + velocity F1 (%)↑ | Pedal activation P (%)↑ | Pedal activation R (%)↑ | Pedal activation F1 (%)↑ | Pedal onset P (%)↑ | Pedal onset R (%)↑ | Pedal onset F1 (%)↑ | Pedal onset+offset P (%)↑ | Pedal onset+offset R (%)↑ | Pedal onset+offset F1 (%)↑ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | TransKun V2 AugPublic weights · Inference code | MIREX 2024 published row | 93.89 | 95.18 | 94.48 | 99.7 | 98.01 | 98.84 | 92.84 | 91.28 | 92.05 | 89.74 | 88.23 | 88.97 | 94.91 | 94.28 | 94.47 | 87.88 | 80.43 | 83.83 | 82.08 | 75.26 | 78.37 |
| 2 | TransKun V2Public weights · Inference code | MIREX 2024 published row | 92.03 | 94.91 | 93.4 | 98.16 | 97.66 | 97.9 | 90.13 | 89.68 | 89.9 | 82.55 | 82.11 | 82.32 | 93.64 | 95.07 | 94.2 | 87.22 | 81.01 | 83.88 | 80.3 | 74.71 | 77.3 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 9 / source rows 9
Scores come from the linked base checkpoint, not a new measurement of the model running here.
| Metric rank | Model | Test setting | Activation | Note Onset | Note Onset+Offset | Note Onset+Offset+ vel. | pedal activation | pedal onset | pedal onset+offset | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | prec↑ | recall↑ | F1↑ | |||
| 1 | Transkun V2Public weights · Inference code | Maestro V3 | 0.9576 | 0.9489 | 0.953 | 0.9956 | 0.9714 | 0.9832 | 0.9465 | 0.9238 | 0.9349 | 0.9411 | 0.9186 | 0.9296 | 0.9671 | 0.9453 | 0.9541 | 0.8909 | 0.8421 | 0.8642 | 0.8632 | 0.8165 | 0.8377 |
| 2 | Transkun V2 AugPublic weights · Inference code | Maestro V3 | 0.9495 | 0.9522 | 0.9505 | 0.9971 | 0.9715 | 0.984 | 0.9437 | 0.9197 | 0.9314 | 0.9386 | 0.9149 | 0.9264 | 0.9546 | 0.9416 | 0.9454 | 0.8883 | 0.8116 | 0.8453 | 0.8497 | 0.7790 | 0.8102 |
| - | Transkun V2 No ExtPublic weights · Inference code | Maestro V3 No Ext | 0.8671 | 0.825 | 0.8441 | 0.9984 | 0.9691 | 0.9833 | 0.8271 | 0.8034 | 0.8149 | 0.823 | 0.7995 | 0.8109 | 0.9498 | 0.9518 | 0.9487 | 0.8872 | 0.8105 | 0.8444 | 0.8413 | 0.7723 | 0.8031 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 2 / source rows 7
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
HPPNet-sp · HFT-transformer · Yan et al. · PAR · DiffRoll (k=9)
Values come from the linked public source. Row labels identify the test set or checkpoint; unreported scores are shown as -.
| Metric rank | Model | # Param | Note Metrics | Note F1 w\offsets (%) | ||||
|---|---|---|---|---|---|---|---|---|
| P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1(%)↑ | |||
| 1 | D3RM (Proposed)Public weights · Inference code | 2.0M | 98.72 | 96.48 | 97.57 | 91.48 | 89.44 | 90.44 |
| 2 | BaselinePublic weights · Inference code | 1.5M | 99.08 | 94.03 | 96.45 | 89.93 | 85.45 | 87.60 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 1 / source rows 6
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Onsets & frames · Onsets & frames [reproduced] · Adversarial onsets & frames · Regress cond on. · Regress cond on. & off.
MAESTRO note results from Table I and sustain-pedal results from Table VI, including published frame, onset, offset and velocity P / R / F1 (%). Unreported metrics remain absent.
| Metric rank | Model | Frame | Note | Note w/ offset | Note w/ offset & vel. | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1 (%)↑ | ||
| - | Regress cond on. & off. & vel.Public weights · Inference code | 88.71 | 90.73 | 89.62 | 98.17 | 95.35 | 96.72 | 83.68 | 81.32 | 82.47 | 82.10 | 79.80 | 80.92 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 1 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Liang · Onsets & frames [our-implemented] · Onsets & frames (misaligned labels) · proposed (misaligned labels)
| Metric rank | Model | Frame | Event | Event w/ offset | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1 (%)↑ | P (%)↑ | R (%)↑ | F1 (%)↑ | ||
| - | ProposedPublic weights · Inference code | 94.30 | 94.42 | 94.25 | 91.59 | 92.41 | 91.86 | 86.36 | 87.02 | 86.58 |
Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.
Deployable entries 1 / source rows 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Onsets & Frames · HPPNet-sp · hFT-Transformer · Transkun
MAESTRO v3 piano evaluation retains onset and onset-offset P/R/F1. The released piano model uses fixed velocity 100 and does not predict pedals.
| Metric rank | Model | Params | Onset | Onset & Offset | ||||
|---|---|---|---|---|---|---|---|---|
| P (%)↑ | R (%)↑ | F1(%)↑ | P (%)↑ | R (%)↑ | F1(%)↑ | |||
| - | SFT-CRNNPublic weights · Inference code | 15M | 99.16 | 97.46 | 98.36 | 91.56 | 90.02 | 90.78 |
Find processing methods and steps suited to your material.