MuScriptor: An Open Model for Multi-Instrument Music Transcription
arXiv - 2026
Transcribe music into multi-instrument MIDI with MuScriptor-large.
Loading...
Turn songs or ensembles into multitrack MIDI to correct notes, organize parts and edit arrangements in a DAW.
Input
Supports inputs in formats such as mp3 / wav / flac / ogg / m4a / aac / aif / aiff / mp4 / mov / mkv
Supported formats
Output
MIDI
Use case
Edit notes, instrumentation and performance details.
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.
Default segment repair reduces instrument errors carrying across segments while preserving sustain; start here for full-mix MIDI drafts.
YourMT3+ · YPTF.MoE+Multi noPS: Editable MIDI drafts from full mixes, plus cleaner solo/common-instrument inputs such as piano/keyboard, guitar, bass, drums, strings/winds, and vocal drafts
YMT3+: Editable MIDI drafts from full mixes, plus cleaner solo/common-instrument inputs such as piano/keyboard, guitar, bass, drums, strings/winds, and vocal drafts
YourMT3+ · YPTF.Single noPS: Editable MIDI drafts from full mixes, plus cleaner solo/common-instrument inputs such as piano/keyboard, guitar, bass, drums, strings/winds, and vocal drafts
YourMT3+ · YPTF.Multi PS: Editable MIDI drafts from full mixes, plus cleaner solo/common-instrument inputs such as piano/keyboard, guitar, bass, drums, strings/winds, and vocal drafts
YourMT3+ · YPTF.MoE+Multi PS: Editable MIDI drafts from full mixes, plus cleaner solo/common-instrument inputs such as piano/keyboard, guitar, bass, drums, strings/winds, and vocal drafts
MIROS MusicFM multi-decoder runs the official local MIROS route with pretrained_msd.pt plus the fine-tuned last.ckpt checkpoint.
BS-RoFormer SW Fixed uses BS-Rofo-SW-Fixed.ckpt plus BS-Rofo-SW-Fixed.yaml to split a full mix into six WAV stems.
OaFS detects 34 instrument classes and exports MIDI and note tables with fixed velocity 127.
Group notes into a specified number of unnamed timbre tracks. Instrument names are not identified. Best suited to short ensembles with a known instrument count; overlapping parts at the same pitch may be inseparable.
Separate vocals, drums, bass, guitar, piano and other sounds, then transcribe and merge MIDI. Velocity prediction and instrument reassignment are optional; separation errors can affect notes.
Separate six stems, then transcribe MIDI. SheetSage2-Refined detects beats, meter and sections; Tsumugi detects chords and key. The existing full-song transcription remains selectable.
arXiv - 2026
Transcribe music into multi-instrument MIDI with MuScriptor-large.
IEEE MLSP - 2024
YourMT3+ extends MT3 with enhanced transformer architecture, MoE variants, multi-channel decoding, and cross-dataset stem augmentation.
ICLR - 2022
MT3 established multi-task multitrack music transcription with sequence modeling and remains the named baseline for newer AMT comparisons.
arXiv / OpenReview - 2026
The 2025 AMT Challenge summary paper was released on 2026-03-29 for the January 1-April 30, 2025 competition. It reports 21 registered teams, 14 submitting teams, 8 valid scored teams, and a 76-piece classical test set with up to three instruments per excerpt.
Signals - 2026
OaFS detects 34 instrument classes and exports MIDI and note tables with fixed velocity 127.
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.
The selected model has no verified matching public result and has not been ranked.
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 3 / 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.
𝒟 Synth · MuScriptor trained on · 1 · 𝒟 Synth · MuScriptor trained on · 2 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 1 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 2
On the same private D_Test, the paper compares the final CFG=1 model with YourMT3+. The results are not independently reproduced and cannot be compared directly with MIROS or other evaluation protocols.
| Metric rank | Model | Test setting | F1 Score ( ↑ ) | ||||
|---|---|---|---|---|---|---|---|
| Onset | Frame | Offset | Drums | Multi | |||
| 1 | MuScriptor (1.3B) · 𝒟 Synth + 𝒟 Real + 𝒟 RLPublic weights · Inference codemodel.safetensorsInference codeWeight license (repository declaration): cc-by-nc-4.0Terms acceptance or access request required | D_Test · CFG = 1 | 60.4 | 73.3 | 49.0 | 50.2 | 48.2 |
| 2 | YourMT3+ (YPTF.MoE+Multi (noPS)) | D_Test · CFG = 1 | 32.52 | 45.54 | 17.79 | 41.4 | 21.9 |
Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.
Deployable entries 3 / 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.
𝒟 Synth · MuScriptor trained on · 1 · 𝒟 Synth · MuScriptor trained on · 2 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 1 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 2
Deployable entries 16 / source rows 16
Deployable entries 0 / source rows 4
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
MuScriptor (60M) · MuScriptor (100M) · MuScriptor (300M) · MuScriptor (1.3B)
Deployable entries 9 / source rows 17
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
GIU · press_to_win · press_to_win5 · press_to_win3 · press_to_win4 · Bytedance Piano transcription · press_to_win2 · Midnight Questor
Deployable entries 9 / source rows 14
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
press_to_win 1 · press_to_win 2 · press_to_win 3 · Bytedance Piano · press_to_win 4
Deployable entries 9 / source rows 9
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.
MT3 (our colab) · MT3 (colab ) · MT3 · PerceiverTF
Deployable entries 1 / 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.
YMT3 · noPS · YMT3+ · noPS | PS · YPTF+S · noPS | PS · YPTF+M · noPS | PS · MT3 [ 1 ] · (colab) · AMT · Baseline *
Deployable entries 1 / source rows 12
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
YMT3 base · + Intra-aug. · + Full-vocab. · + Data balancing · + Cross-aug. · + PTF-encoder · + FFN → MoE · + Multi decoder · MT3 (colab) · MT3 [ 1 ] · PerceiverTF [ 25 ]
Deployable entries 10 / source rows 25
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
OaFS-ResUnet · Frame F1 (%) · OaFS-CFP · Frame F1 (%) · OaFS-single · Frame F1 (%) · OaFS-ResUnet · Onset F1 (%) · OaFS-CFP · Onset F1 (%) · OaFS-single · Onset F1 (%) · OaFS-ResUnet · Multi-Instrument Frame F1 (%) · OaFS-CFP · Multi-Instrument Frame F1 (%) · OaFS-single · Multi-Instrument Frame F1 (%) · OaFS-ResUnet · Multi-Instrument Onset F1 (%) · OaFS-CFP · Multi-Instrument Onset F1 (%) · OaFS-single · Multi-Instrument Onset F1 (%) · OaFS-ResUnet · Instrument Detection F1 (%) · OaFS-CFP · Instrument Detection F1 (%) · OaFS-single · Instrument Detection F1 (%)
Deployable entries 68 / source rows 102
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
OaFS-Single · Acoustic Piano · OaFS-Single · Electric Piano · OaFS-Single · Chromatic Percussion · OaFS-Single · Organ · OaFS-Single · Acoustic Guitar · OaFS-Single · Clean Electric Guitar · OaFS-Single · Distorted Electric Guitar · OaFS-Single · Acoustic Bass · OaFS-Single · Electric Bass · OaFS-Single · Violin · OaFS-Single · Viola · OaFS-Single · Cello · OaFS-Single · Contrabass · OaFS-Single · Orchestral Harp · OaFS-Single · Timpani · OaFS-Single · String Ensemble · OaFS-Single · Synth Strings · OaFS-Single · Choir and Voice · OaFS-Single · Trumpet · OaFS-Single · Trombone · OaFS-Single · Tuba · OaFS-Single · French Horn · OaFS-Single · Brass Section · OaFS-Single · Soprano/Alto Sax · OaFS-Single · Tenor Sax · OaFS-Single · Baritone Sax · OaFS-Single · Oboe · OaFS-Single · English Horn · OaFS-Single · Bassoon · OaFS-Single · Clarinet · OaFS-Single · Pipe · OaFS-Single · Synth Lead · OaFS-Single · Synth Pad · OaFS-Single · Drums
Deployable entries 28 / source rows 28
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.
OaFS-ResUnet · OaFS-CFP · OaFS-single
Deployable entries 3 / 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.
D16 · MSE · Syn · Rescale
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 -.
Deployable entries 3 / 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.
𝒟 Synth · MuScriptor trained on · 1 · 𝒟 Synth · MuScriptor trained on · 2 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 1 · 𝒟 Synth + 𝒟 Real · MuScriptor trained on · 2
On the same private D_Test, the paper compares the final CFG=1 model with YourMT3+. The results are not independently reproduced and cannot be compared directly with MIROS or other evaluation protocols.
| Metric rank | Model | Test setting | F1 Score ( ↑ ) | ||||
|---|---|---|---|---|---|---|---|
| Onset | Frame | Offset | Drums | Multi | |||
| 1 | MuScriptor (1.3B) · 𝒟 Synth + 𝒟 Real + 𝒟 RLPublic weights · Inference codemodel.safetensorsInference codeWeight license (repository declaration): cc-by-nc-4.0Terms acceptance or access request required | D_Test · CFG = 1 | 60.4 | 73.3 | 49.0 | 50.2 | 48.2 |
| 2 | YourMT3+ (YPTF.MoE+Multi (noPS)) | D_Test · CFG = 1 | 32.52 | 45.54 | 17.79 | 41.4 | 21.9 |
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 -.
Deployable entries 16 / source rows 16
Paper Table 2 evaluates the two released, locally deployable models on eight datasets excluded from both training sets.
| Metric rank | Model | Test setting | F1 Score ( ↑ ) | ||||
|---|---|---|---|---|---|---|---|
| Onset | Frame | Offset | Drums | Multi | |||
| 1 | MuScriptorPublic weights · Inference codemodel.safetensorsInference codeWeight license (repository declaration): cc-by-nc-4.0Terms acceptance or access request required | Bach10 | 43.1 | 85.0 | 36.0 | - | 34.7 |
| 2 | YourMT3+ | Bach10 | 59.8 | 66.0 | 48.0 | - | 26.4 |
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 -.
Deployable entries 9 / source rows 17
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
GIU · press_to_win · press_to_win5 · press_to_win3 · press_to_win4 · Bytedance Piano transcription · press_to_win2 · Midnight Questor
Reproducible open-source entries from the official leaderboard, with official ranks preserved. F1, precision, recall, onset, offset and overlap use percentages; runtime is in milliseconds. The YourMT3+ competition recipe uses instrument information supplied in filenames; its score is not an end-to-end measurement of this site’s processing pipeline.
| Official rank | Model | Members | Submission date | Runtime (ms) | F-measure (summary)↑ | Precision↑ | Recall↑ | Average overlap ratio↑ | F-measure (details)↑ | Onset precision↑ | Onset recall↑ | Onset F-measure↑ | Offset precision↑ | Offset recall↑ | Offset F-measure↑ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | MIROS (Music Information Retrieval Osnabrück)Public weights · Inference codeMIROS fine-tuned decoderMusicFM encoderInference code logs/Multi_longer_seq_length_frozen_enc_silu/le2bzt53/checkpoints/last.ckptEvaluation settingsWeight license not declared | Deniz Gün, Fernando Riveros, Paul Koesling, Anish Haluvani Sundresh, William Shelor | 2025-06-02 | 22.0535 | 0.5999 | 0.655839 | 0.5724 | 0.739079 | 0.599875 | 0.877975 | 0.759886 | 0.799196 | 0.807888 | 0.701962 | 0.736779 |
| 2 | YourMT3-YPTF-MoE-M | Sungkyun Chang, Simon Dixon, Emmanouil Benetos | 2025-06-03 | 12.5981 | 0.5939 | 0.600966 | 0.588778 | 0.730518 | 0.593868 | 0.792573 | 0.777088 | 0.783356 | 0.767444 | 0.750847 | 0.757675 |
| 3 | YourMT3-YPTF-S | Sungkyun Chang, Simon Dixon, Emmanouil Benetos | 2025-06-02 | 15.3954 | 0.5582 | 0.556491 | 0.561459 | 0.732617 | 0.55816 | 0.762297 | 0.771807 | 0.76566 | 0.743591 | 0.753287 | 0.747136 |
| 4 | YourMT3-P | Sungkyun Chang, Simon Dixon, Emmanouil Benetos | 2025-06-02 | 14.9916 | 0.3948 | 0.39658 | 0.398548 | 0.72627 | 0.394752 | 0.60259 | 0.613207 | 0.601815 | 0.685018 | 0.691227 | 0.681928 |
| 5 | MT3Public weights · Inference code | Google Research Brain Team: Josh Gardner, Ian Simon, Ethan Manilow, Curtis Hawthorne, Jesse Engel | 2025-06-02 | 20.1915 | 0.3932 | 0.38112 | 0.411528 | 0.71797 | 0.393214 | 0.663145 | 0.726103 | 0.687734 | 0.63501 | 0.698339 | 0.659656 |
| 7 | YourMT3-YPTF-SP-VPublic weights · Inference codelogs/2024/ptfs_gmlp_wf_e12_d8_500k/checkpoints/epoch=466-step=500000.ckptInference codeEvaluation settingsWeight license not declared | Sungkyun Chang, Simon Dixon, Emmanouil Benetos | 2025-06-02 | 14.5031 | 0.3306 | 0.32796 | 0.335827 | 0.714713 | 0.330591 | 0.632511 | 0.657221 | 0.641741 | 0.686195 | 0.711273 | 0.695724 |
| 10 | YourMT3-YPTF-MoE-MP | Sungkyun Chang, Simon Dixon, Emmanouil Benetos | 2025-06-02 | 16.0262 | 0.2173 | 0.214966 | 0.220641 | 0.611634 | 0.217338 | 0.696075 | 0.719896 | 0.70555 | 0.699835 | 0.72405 | 0.709509 |
| 15 | ReconVATPublic weights · Inference code | Kin Wai Cheuk, Dorien Herremans, Li Su | 2025-06-02 | 5.4548 | 0.1416 | 0.121534 | 0.180252 | 0.789788 | 0.141598 | 0.307451 | 0.450635 | 0.35515 | 0.478656 | 0.742044 | 0.564281 |
| 16 | Basic PitchPublic weights · Inference code | Spotify's Audio Intelligence Lab: Rachel M. Bittner, Juan José Bosch, David Rubinstein, Gabriel Meseguer-Brocal, Sebastian Ewert | 2025-06-02 | 3.9056 | 0.0634 | 0.054995 | 0.078152 | 0.597734 | 0.063433 | 0.264086 | 0.337363 | 0.290772 | 0.444662 | 0.587456 | 0.495507 |
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 -.
Deployable entries 9 / source rows 14
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
press_to_win 1 · press_to_win 2 · press_to_win 3 · Bytedance Piano · press_to_win 4
The website and paper are separate snapshots with small last-digit differences. The test contains 76 constrained synthetic excerpts, with up to three instruments each. Values follow their respective sources. The YourMT3+ competition recipe uses instrument information supplied in filenames; its score is not an end-to-end measurement of this site’s processing pipeline.
| Official rank | Model | F1 Score↑ | Precision↑ | Recall↑ | Overlap↑ | Runtime |
|---|---|---|---|---|---|---|
| 1 | MIROSPublic weights · Inference codeMIROS fine-tuned decoderMusicFM encoderInference code logs/Multi_longer_seq_length_frozen_enc_silu/le2bzt53/checkpoints/last.ckptEvaluation settingsWeight license not declared | 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 (baseline)Public weights · Inference code | 0.3932 | 0.3811 | 0.4115 | 0.7180 | 20.19 |
| 6 | YourMT3-YPTF-SP-VPublic weights · Inference codelogs/2024/ptfs_gmlp_wf_e12_d8_500k/checkpoints/epoch=466-step=500000.ckptInference codeEvaluation settingsWeight license not declared | 0.3305 | 0.3280 | 0.3358 | 0.7147 | 14.50 |
| 9 | YourMT3-YPTF-MoE-MP | 0.2173 | 0.2150 | 0.2206 | 0.6116 | 16.03 |
| 13 | ReconVATPublic weights · Inference code | 0.1415 | 0.1215 | 0.1803 | 0.7898 | 5.45 |
| 14 | Basic PitchPublic weights · Inference code | 0.0634 | 0.0550 | 0.0782 | 0.5977 | 3.91 |
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 -.
Deployable entries 9 / source rows 9
Table 2 groups exact instrument counts and reports mean ± standard deviation. Groups are ranked separately and do not replace the overall table. The YourMT3+ competition recipe uses instrument information supplied in filenames; its score is not an end-to-end measurement of this site’s processing pipeline.
| Metric rank | Model | Test setting | F-measure↑ | Precision↑ | Recall↑ |
|---|---|---|---|---|---|
| 1 | YourMT3-YPTF-MoE-M | Count = 1 (n=6) | 0.7594 ± 0.2304 | 0.7858 ± 0.2411 | 0.7367 ± 0.2254 |
| 2 | MIROSPublic weights · Inference codeMIROS fine-tuned decoderMusicFM encoderInference code logs/Multi_longer_seq_length_frozen_enc_silu/le2bzt53/checkpoints/last.ckptEvaluation settingsWeight license not declared | Count = 1 (n=6) | 0.7193 ± 0.2103 | 0.9067 ± 0.2286 | 0.6233 ± 0.2295 |
| 3 | MT3 (Baseline)Public weights · Inference code | Count = 1 (n=6) | 0.5483 ± 0.3898 | 0.5548 ± 0.3974 | 0.5422 ± 0.3829 |
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.
MT3 (our colab) · MT3 (colab ) · MT3 · PerceiverTF
YourMT3+ paper comparison against MT3-family baselines on Slakh-style metrics.
| Metric rank | Model | Bass↑ | Brass↑ | C.Perc↑ | Guitar↑ | Organ↑ | Piano↑ | Pipe↑ | Reed↑ | Strings and Ens.↑ | Synth Lead↑ | Synth Pad↑ | Drums↑ | Onset F1 (non-drum)↑ | Offset F1 (non-drum)↑ | Onset F1 (drums/gm)↑ | Multi (Onset-Offset) F1 ♠↑ | Multi Onset F1↑ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| - | YPTF.MoE+M (PS) | 93.20 | 74.96 | 67.70 | 82.27 | 73.48 | 88.84 | 74.72 | 82.22 | 75.44 | 84.19 | 45.57 | 90.05 | 84.56 | 70.70 | 90.05 | 74.84 | - |
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 7
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
YMT3 · noPS · YMT3+ · noPS | PS · YPTF+S · noPS | PS · YPTF+M · noPS | PS · MT3 [ 1 ] · (colab) · AMT · Baseline *
| Metric rank | Model | Test setting | MAPS [ 21 ] (unseen) · Piano | MAPS [ 21 ] (seen) | Maestro v3 | MusicNet ext. (EM) [ 10 ] | MusicNet ext. [ 9 , 10 ] | MIR-ST500 [ 13 ] (SVS) · Singing | MIR-ST500 [ 13 ] | MIR-ST500 (100ms [ 20 ] ) | ENSTdrums (DTP [ 15 ] ) · Drums | ENSTdrums (DTM [ 15 ] ) | GuitarSet [ 11 ] (MT3 [ 1 ] ) · Guitar | URMP [ 22 ] Onset F1 [ 1 ] · Agnostic | URMP [ 22 ] Multi F1 [ 1 ] · Ensemble | Slakh [ 18 ] Onset F1 [ 1 ] · Agnostic | Slakh [ 18 ] Multi F1 [ 1 ] · All | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Strings↑ | Winds↑ | Strings↑ | Winds↑ | ||||||||||||||||
| - | YPTF.MoE+M | noPS | PS | 87.88 | 86.25 | - | 96.98 | 96.52 | 91.32 | 90.07 | 83.46 | 78.50 | 66.14 | 66.09 | 55.95 | 55.33 | 71.60 | 72.05 | 70.59 | 71.07 | 81.14 | 82.08 | 88.79 | 89.48 | 85.92 | 87.27 | 91.65 | 88.87 | 81.05 | 81.79 | 67.22 | 67.98 | 84.14 | 84.56 | 73.98 | 74.84 |
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 12
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
YMT3 base · + Intra-aug. · + Full-vocab. · + Data balancing · + Cross-aug. · + PTF-encoder · + FFN → MoE · + Multi decoder · MT3 (colab) · MT3 [ 1 ] · PerceiverTF [ 25 ]
| Metric rank | Model | Onset F1↑ | Offset F1↑ | Drum F1↑ |
|---|---|---|---|---|
| - | YPTF.MoE+Multi | 84.6 | 70.7 | 90.1 |
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 10 / source rows 25
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
OaFS-ResUnet · Frame F1 (%) · OaFS-CFP · Frame F1 (%) · OaFS-single · Frame F1 (%) · OaFS-ResUnet · Onset F1 (%) · OaFS-CFP · Onset F1 (%) · OaFS-single · Onset F1 (%) · OaFS-ResUnet · Multi-Instrument Frame F1 (%) · OaFS-CFP · Multi-Instrument Frame F1 (%) · OaFS-single · Multi-Instrument Frame F1 (%) · OaFS-ResUnet · Multi-Instrument Onset F1 (%) · OaFS-CFP · Multi-Instrument Onset F1 (%) · OaFS-single · Multi-Instrument Onset F1 (%) · OaFS-ResUnet · Instrument Detection F1 (%) · OaFS-CFP · Instrument Detection F1 (%) · OaFS-single · Instrument Detection F1 (%)
Official Table 5 compares five datasets separately and shows only experiments with verified matching public weights. The current OaFS is jointly trained; these results do not establish superiority over MuScriptor.
| Metric rank | Model | Test setting | Frame F1 (%)↑ | Onset F1 (%)↑ | Multi-Instrument Frame F1 (%)↑ | Multi-Instrument Onset F1 (%)↑ | Instrument Detection F1 (%)↑ |
|---|---|---|---|---|---|---|---|
| 1 | MT3Public weights · Inference codeOfficial MT3 T5X checkpoint and 294 tensor filesInference codeEvaluation settingsWeight license not declared | MAESTRO | 55.30 | 95.16 | 55.30 | 95.16 | 99.72 |
| 2 | OaFSPublic weights · Inference code | MAESTRO | 76.10 | 94.17 | 76.10 | 94.17 | 100 |
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 68 / source rows 102
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
OaFS-Single · Acoustic Piano · OaFS-Single · Electric Piano · OaFS-Single · Chromatic Percussion · OaFS-Single · Organ · OaFS-Single · Acoustic Guitar · OaFS-Single · Clean Electric Guitar · OaFS-Single · Distorted Electric Guitar · OaFS-Single · Acoustic Bass · OaFS-Single · Electric Bass · OaFS-Single · Violin · OaFS-Single · Viola · OaFS-Single · Cello · OaFS-Single · Contrabass · OaFS-Single · Orchestral Harp · OaFS-Single · Timpani · OaFS-Single · String Ensemble · OaFS-Single · Synth Strings · OaFS-Single · Choir and Voice · OaFS-Single · Trumpet · OaFS-Single · Trombone · OaFS-Single · Tuba · OaFS-Single · French Horn · OaFS-Single · Brass Section · OaFS-Single · Soprano/Alto Sax · OaFS-Single · Tenor Sax · OaFS-Single · Baritone Sax · OaFS-Single · Oboe · OaFS-Single · English Horn · OaFS-Single · Bassoon · OaFS-Single · Clarinet · OaFS-Single · Pipe · OaFS-Single · Synth Lead · OaFS-Single · Synth Pad · OaFS-Single · Drums
| Metric rank | Model | Test setting | Total Number of Notes | Frame F1 (%)↑ | Onset F1 (%)↑ |
|---|---|---|---|---|---|
| 1 | OaFSPublic weights · Inference code | Acoustic Piano | 974,080 | 65.90 | 71.37 |
| 2 | MT3Public weights · Inference codeOfficial MT3 T5X checkpoint and 294 tensor filesInference codeEvaluation settingsWeight license not declared | Acoustic Piano | 974,080 | 56.90 | 66.08 |
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 28 / source rows 28
| Metric rank | Model | Test setting | Frame F1 (%)↑ | Onset F1 (%)↑ |
|---|---|---|---|---|
| 1 | OaFSPublic weights · Inference code | Piano | 70.36 | 72.10 |
| 2 | MT3Public weights · Inference codeOfficial MT3 T5X checkpoint and 294 tensor filesInference codeEvaluation settingsWeight license not declared | Piano | 63.12 | 69.99 |
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 5
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
OaFS-ResUnet · OaFS-CFP · OaFS-single
| Metric rank | Model | Number of Trainable Parameters | Number of Instrument Classes Excluding “Others” |
|---|---|---|---|
| - | MT3Public weights · Inference codeOfficial MT3 T5X checkpoint and 294 tensor filesInference codeEvaluation settingsWeight license not declared | 93.7 M | 34 |
| - | OaFSPublic weights · Inference code | 64.9 M | 34 |
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 3 / 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.
D16 · MSE · Syn · Rescale
Paper Table 3: K is the specified timbre count. F1 is computed per track after optimal track matching. Results average at least three training runs; they are not a fresh measurement of this site's checkpoint.
| Metric rank | Model | BACH10 2 mix | BACH10 3 mix | BACH10 4 mix | URMP 2 mix | URMP 3 mix | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| F FS↑ | F S↑ | ratio↑ | F FS↑ | F S↑ | ratio↑ | F FS↑ | F S↑ | ratio↑ | F FS↑ | F S↑ | ratio↑ | F FS↑ | F S↑ | ratio↑ | ||
| 1 | OursPublic weights · Inference code | 83.4 | 84.6 | 99.0 | 77.8 | 80.1 | 96.7 | 66.2 | 72.7 | 89.9 | 68.9 | 66.8 | 82.6 | 58.5 | 60.0 | 77.1 |
| 2 | SharePublic weights · Inference codeseptimbre/ablation/DicephNet/dicephnet_model.pthInference codeEvaluation settingsWeight license not declared | 82.1 | 83.0 | 98.4 | 76.1 | 78.9 | 96.6 | 68.5 | 72.8 | 91.0 | 69.0 | 68.3 | 84.6 | 57.2 | 57.5 | 74.1 |
| 3 | TanakaPublic weights · Inference code | 77.9 | 79.6 | 93.2 | 66.1 | 69.0 | 83.4 | 55.4 | 59.2 | 73.2 | 65.5 | 64.1 | 79.3 | 56.5 | 56.5 | 72.8 |
Find processing methods and steps suited to your material.
TelkNet fix note · August 13, 2026
We ran the same full song through TelkNet and official MuScriptor Large 0.3.0 in strict mode. Both produced 8 parts and 6,560 notes, with the same part, pitch, and order for every note. The anomaly sits around 182–222 seconds: the source still has a clear pitched line in the mid-to-high frequencies, but the MIDI has a roughly 40.31-second gap above C4. The UI, BPM processing, and MIDI editor did not remove those notes; the gap is already in the official model output. Without the original score, we cannot say exactly how many notes are missing.
TelkNet adjusts how instrument information passes between 5-second segments to reduce the spread of misclassifications and connect sustained notes across segments. MuScriptor Large transcribes each segment once.
We then completed 16 full runs across 14 songs. On the first song, the longest onset gap above C4 fell from 40.309 seconds to 1.970 seconds. On the second, the output changed from 1,698 notes on 6 active tracks with 70.79% piano to 2,279 notes on 8 active tracks with 50.77% piano. Two full reruns of each song produced byte-identical files.
MuScriptor tempo compatibility
The official baseline uses Beat This final0 and a free-intercept least-squares fit to publish one constant BPM. Strict mode fails when the detected beats do not support a constant tempo; best-effort mode writes a placeholder 120 BPM. The MIDI writer also corrects onset delay and shifts all notes forward to align a bar line with the downbeat.
TelkNet keeps the same Beat This final0 detector, and an automatically detected single BPM remains the recommended default. When the user selects following source-tempo changes, one BPM is used only if the whole-song grid stays within 40 ms of the detected beats; otherwise an adaptive Tempo Map is published. Manual BPM remains selectable. With quantization off, model-native note times remain unchanged.
| Comparison | Official MuScriptor v0.3.0 | TelkNet v11 |
|---|---|---|
| Beat detection | Beat This final0. | The same Beat This final0; the detector model is not replaced. |
| Published tempo grid | A free-intercept fit publishes one constant BPM. Strict mode fails for non-constant tempo; best-effort writes a placeholder 120 BPM. | Fixed BPM only when the maximum grid distance is at most 40 ms; otherwise a phase-bounded adaptive Tempo Map. |
| BPM control | The current official CLI has no manual BPM option. | Adaptive, fixed-auto, and manual BPM from 30 to 300. |
| Note timeline | Corrects onset delay and shifts all notes forward to align a bar line with the downbeat. | With quantization off, model-native note times remain unchanged and no bar-alignment shift is applied. |
Beat This remains an audio estimator. Unusual meter, weak or transient-free rhythm, rubato, abrupt tempo changes, or a wrong half/double-tempo pulse can still produce an incorrect grid. Automatic BPM may differ slightly from a known exact value, so Fixed · Manual BPM is available when the source BPM is known. Enabling note quantization intentionally moves note times. Tempo settings cannot correct missing notes, wrong pitches, or instrument and track-label errors from MuScriptor. Important MIDI should still be checked against the source in a DAW.