Loading...
Convert solo or isolated instrument recordings to MIDI, then edit pitches and timing.
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.
FiloSax targets monophonic jazz saxophone solos and exports discrete pitch, onset, and offset events as MIDI. It does not output continuous pitch, slide, or vibrato tracks.
SwiftF0 0.2.0 uses its official pitch tracking and note segmentation for monophonic recordings. Outputs MIDI, frame-level pitch CSV and a plot.
SwiftF0 0.3.0 tracks pitch in monophonic recordings and exports MIDI, frame-level pitch CSV and charts.
Spotify Basic Pitch transcribes polyphonic performances from nearly any instrument, including voice, into MIDI and optional note-event CSV with pitch bends. It works best on one instrument at a time.
GAPS converts clean or separated solo guitar into note events and MIDI, without string positions, frets, continuous F0 or Pitch Bend. Adjust MIDI velocity by comparing with the original recording.
Guitar-FL converts guitar solos from the François Leduc data domain into discrete note events and MIDI; FL does not mean frame-level. It does not output continuous F0 or Pitch Bend.
FiloBass targets monophonic jazz double bass (upright bass), not general or electric bass. Use a clean or separated stem; kick and low-frequency bleed can create false notes.
VioPTT outputs violin note events and one of four technique labels—détaché, flageolet, spiccato, pizzicato—or none. MIDI velocity is fixed, not measured performance dynamics.
STRAdi Online targets clean solo violin with fully causal inference and no future context. It exports discrete MIDI plus onset, offset and pitch CSV; the public repository does not ship a score-following system.
STRAdi Offline is the matched non-causal, full-context reference route. It exports the same discrete MIDI and onset, offset and pitch CSV; it is not a published score-following system either.
Convert guitar recordings to MIDI with the GAPS paper checkpoint trained for 12200 iterations.
Tsumugi · Bass v2: Edit notes, instrumentation and performance details.
Tsumugi · Guitar v1.5: Edit notes, instrumentation and performance details.
Tsumugi · Other v1.5: Edit notes, instrumentation and performance details.
Tsumugi · Vocal: Edit notes, instrumentation and performance details.
Tsumugi · Vocal Harmony v1.6: Edit notes, instrumentation and performance details.
Tsumugi · Vocal Harmony v1.5: Edit notes, instrumentation and performance details.
Tsumugi · Drums v1.5: Edit notes, instrumentation and performance details.
ISMIR - 2024
GAPS converts clean or separated solo guitar into note events and MIDI, without string positions, frets, continuous F0 or Pitch Bend. Adjust MIDI velocity by comparing with the original recording.
ICASSP - 2024
Guitar-FL converts guitar solos from the François Leduc data domain into discrete note events and MIDI; FL does not mean frame-level. It does not output continuous F0 or Pitch Bend.
SMC - 2024
ISMIR - 2023
FiloBass targets monophonic jazz double bass (upright bass), not general or electric bass. Use a clean or separated stem; kick and low-frequency bleed can create false notes.
ICASSP - 2026
VioPTT outputs violin note events and one of four technique labels—détaché, flageolet, spiccato, pizzicato—or none. MIDI velocity is fixed, not measured performance dynamics.
arXiv - 2025
SwiftF0 0.2.0 uses its official pitch tracking and note segmentation for monophonic recordings. Outputs MIDI, frame-level pitch CSV and a plot.
ICASSP - 2022
Spotify Basic Pitch transcribes polyphonic performances from nearly any instrument, including voice, into MIDI and optional note-event CSV with pitch bends. It works best on one instrument at a time.
Official technical guide - 2026
Convert isolated instrument, vocal or drum tracks to MIDI using the matching official model. Vocal harmony also supports the v1.6 checkpoint. Instrument label refinement remains experimental.
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 -.
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
TorchCREPE · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · PESTO · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · HarmoF0 · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · SPICE · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · PENN · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents
SwiftF0 0.2.0 uses the 2026-09-19 snapshot of the author's pitch-benchmark: 10 corpora, 8 scored conditions and 50 cents tolerance. This is pitch F1; the difference from RMVPE is statistically undetermined. Asymmetric 95% CIs are retained. REAPER is unranked due to incomplete results; the older 0.1.x paper does not score 0.2.0.
| Metric rank | Model | Test setting | Pitch F1@50c↑ | Voicing F1↑ | CI95 lo | CI95 hi |
|---|---|---|---|---|---|---|
| 1 | SwiftF0 0.2.0Public weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.778 | 0.839 | 0.765 | 0.792 |
| 2 | RMVPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.768 | 0.837 | 0.752 | 0.783 |
| 3 | FCPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.728 | 0.812 | 0.712 | 0.742 |
| 4 | CREPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.689 | 0.781 | 0.672 | 0.704 |
| 5 | SHSAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.657 | 0.773 | 0.64 | 0.672 |
| 6 | PraatAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.651 | 0.815 | 0.634 | 0.667 |
| 7 | RAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.64 | 0.799 | 0.621 | 0.655 |
| 8 | SWIPEAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.61 | 0.761 | 0.589 | 0.624 |
| 9 | HarvestAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.6 | 0.782 | 0.584 | 0.614 |
| 10 | YAAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.56 | 0.842 | 0.539 | 0.58 |
| 10 | DIOAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.56 | 0.71 | 0.541 | 0.577 |
| 12 | Basic PitchPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.557 | 0.775 | 0.539 | 0.573 |
| 13 | pYINAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.506 | 0.656 | 0.482 | 0.525 |
Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.
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.
Ours w/o aug · Ours + FT w/o aug · Ours + FT w/ aug · MUSC · MERTech
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.
MT3 · Kong et al. · Kong et al. (w/ aug) · Zang et al. (Synthtab) · Maman (MusicNet EM ) · Maman (Guitar) · Riley et al. · Ours · Ours (Finetuned from )
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.
Omnizart · MT3 · Lu et al. · Riley et al. · Ours (Finetuned from )
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.
Riley et al. · Ours · Ours (Finetuned from )
Deployable entries 1 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Başaran et al. · Ours
Deployable entries 1 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Başaran et al. · Ours
Deployable entries 2 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Melodyne
Deployable entries 13 / source rows 18
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
TorchCREPE · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · PESTO · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · HarmoF0 · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · SPICE · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents · PENN · pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents
Deployable entries 13 / source rows 18
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
TorchCREPE · pitch-benchmark · 10 × 8 · 50 cents · PESTO · pitch-benchmark · 10 × 8 · 50 cents · HarmoF0 · pitch-benchmark · 10 × 8 · 50 cents · SPICE · pitch-benchmark · 10 × 8 · 50 cents · PENN · pitch-benchmark · 10 × 8 · 50 cents
Deployable entries 2 / source rows 2
Deployable entries 2 / source rows 2
Deployable entries 2 / source rows 2
Deployable entries 1 / source rows 2
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Base instrument classifier · Held-out RWC-I
Deployable entries 5 / source rows 5
Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.
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.
Ours w/o aug · Ours + FT w/o aug · Ours + FT w/ aug · MUSC · MERTech
The sources use different training or task settings, so compare numbers only within the stated protocol.
| Metric rank | Model | URMP | Bach10 | ||||||
|---|---|---|---|---|---|---|---|---|---|
| P↑ | R↑ | F1↑ | F1 no↑ | P↑ | R↑ | F1↑ | F1 no↑ | ||
| - | Ours w/ augPublic weights · Inference code | 86.1 | 83.6 | 84.5 | 93.1 | 68.1 | 71.8 | 69.9 | 79.5 |
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.
Omnizart · MT3 · Lu et al. · Riley et al. · Ours (Finetuned from )
Paper results retain their datasets. FiloBass Table 5 evaluates other algorithms on that dataset; the current FiloBass checkpoint has no published score.
| Metric rank | Model | P 50↑ | R 50↑ | F 50↑ |
|---|---|---|---|---|
| - | Basic PitchPublic weights · Inference code | 54.6 | 85.0 | 66.1 |
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 1 / source rows 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Başaran et al. · Ours
Paper results retain their datasets. FiloBass Table 5 evaluates other algorithms on that dataset; the current FiloBass checkpoint has no published score.
| Metric rank | Model | P 50↑ | R 50↑ | F 50↑ |
|---|---|---|---|---|
| - | CREPE NotesPublic weights · Inference code | 82.35 | 83.05 | 82.67 |
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 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Başaran et al. · Ours
Paper results retain their datasets. FiloBass Table 5 evaluates other algorithms on that dataset; the current FiloBass checkpoint has no published score.
| Metric rank | Model | P 50↑ | R 50↑ | F 50↑ |
|---|---|---|---|---|
| - | CREPE NotesPublic weights · Inference code | 70.93 | 70.63 | 70.41 |
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 3
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
Melodyne
Paper results retain their datasets. FiloBass Table 5 evaluates other algorithms on that dataset; the current FiloBass checkpoint has no published score.
| Metric rank | Model | R no↑ | P no↑ | F no↑ | O↑ |
|---|---|---|---|---|---|
| 1 | CREPE NotesPublic weights · Inference code | 74.11 ± 12.09 | 71.81 ± 13.33 | 72.89 ± 12.68 | 78.77 ± 2.68 |
| 2 | Basic PitchPublic weights · Inference code | 81.28 ± 6.26 | 51.40 ± 6.28 | 62.73 ± 5.55 | 65.24 ± 4.51 |
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 -.
SwiftF0 0.2.0 uses the 2026-09-19 snapshot of the author's pitch-benchmark: 10 corpora, 8 scored conditions and 50 cents tolerance. This is pitch F1; the difference from RMVPE is statistically undetermined. Asymmetric 95% CIs are retained. REAPER is unranked due to incomplete results; the older 0.1.x paper does not score 0.2.0.
| Metric rank | Model | Test setting | Pitch F1@50c↑ | Voicing F1↑ | CI95 lo | CI95 hi |
|---|---|---|---|---|---|---|
| 1 | SwiftF0 0.2.0Public weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.778 | 0.839 | 0.765 | 0.792 |
| 2 | RMVPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.768 | 0.837 | 0.752 | 0.783 |
| 3 | FCPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.728 | 0.812 | 0.712 | 0.742 |
| 4 | CREPEPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.689 | 0.781 | 0.672 | 0.704 |
| 5 | SHSAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.657 | 0.773 | 0.64 | 0.672 |
| 6 | PraatAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.651 | 0.815 | 0.634 | 0.667 |
| 7 | RAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.64 | 0.799 | 0.621 | 0.655 |
| 8 | SWIPEAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.61 | 0.761 | 0.589 | 0.624 |
| 9 | HarvestAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.6 | 0.782 | 0.584 | 0.614 |
| 10 | YAAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.56 | 0.842 | 0.539 | 0.58 |
| 10 | DIOAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.56 | 0.71 | 0.541 | 0.577 |
| 12 | Basic PitchPublic weights · Inference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.557 | 0.775 | 0.539 | 0.573 |
| 13 | pYINAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 corpora · 8 altered conditions · 50 cents | 0.506 | 0.656 | 0.482 | 0.525 |
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 -.
Author pitch-benchmark revision 6f21369: 10 corpora, 8 altered conditions, 50 cents, frame-level pitch F1. SwiftF0 versus RMVPE is not statistically resolved; this is not MIDI note F1.
| Metric rank | Model | Test setting | Pitch F1@50c↑ | Voicing F1↑ | CI95 lo | CI95 hi |
|---|---|---|---|---|---|---|
| 1 | SwiftF0 0.3.0Public weights · Inference code | pitch-benchmark · 10 × 8 · 50 cents | 0.781 | 0.844 | 0.768 | 0.795 |
| 2 | RMVPEPublic weights · Inference code | pitch-benchmark · 10 × 8 · 50 cents | 0.768 | 0.837 | 0.752 | 0.783 |
| 3 | FCPEPublic weights · Inference code | pitch-benchmark · 10 × 8 · 50 cents | 0.728 | 0.812 | 0.712 | 0.742 |
| 4 | CREPEPublic weights · Inference code | pitch-benchmark · 10 × 8 · 50 cents | 0.689 | 0.781 | 0.672 | 0.704 |
| 5 | SHSAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.657 | 0.773 | 0.64 | 0.672 |
| 6 | PraatAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.651 | 0.815 | 0.634 | 0.667 |
| 7 | RAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.64 | 0.799 | 0.621 | 0.655 |
| 8 | SWIPEAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.61 | 0.761 | 0.589 | 0.624 |
| 9 | HarvestAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.6 | 0.782 | 0.584 | 0.614 |
| 10 | YAAPTAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.56 | 0.842 | 0.539 | 0.58 |
| 10 | DIOAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.56 | 0.71 | 0.541 | 0.577 |
| 12 | BasicPitchPublic weights · Inference code | pitch-benchmark · 10 × 8 · 50 cents | 0.557 | 0.775 | 0.539 | 0.573 |
| 13 | pYINAlgorithm needs no trained weights · Inference codeAlgorithm needs no trained weightsInference code | pitch-benchmark · 10 × 8 · 50 cents | 0.506 | 0.656 | 0.482 | 0.525 |
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 -.
Scores come from the linked base checkpoint, not a new measurement of the model running here.
| Metric rank | Model | Test setting | F_no-offset↑ | Frame accuracy ↑ | F_with-offset ↑ |
|---|---|---|---|---|---|
| - | NMPPublic weights · Inference code | Molina | 0.52 | 0.63 | 0.35 |
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 -.
Tsumugi Other and Vocal Harmony report COnP, not COnF1. Published results cover v1.5; default Vocal Harmony v1.6 has no published score. Instrument refinement remains experimental, with datasets shown separately.
| Metric rank | Model | Test setting | COnP (%)↑ |
|---|---|---|---|
| 1 | other_v1_5Public weights · Inference code | In-house real audio | 77.01 |
| 2 | otherPublic weights · Inference code | In-house real audio | 73.18 |
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 -.
Tsumugi Other and Vocal Harmony report COnP, not COnF1. Published results cover v1.5; default Vocal Harmony v1.6 has no published score. Instrument refinement remains experimental, with datasets shown separately.
| Metric rank | Model | Test setting | COnP (%)↑ |
|---|---|---|---|
| 1 | vocal_harmony_v1_5Public weights · Inference code | Held-out MIR-ST500 | 68.14 |
| 2 | vocal_harmonyPublic weights · Inference code | Held-out MIR-ST500 | 60.52 |
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 -.
Tsumugi Other and Vocal Harmony report COnP, not COnF1. Published results cover v1.5; default Vocal Harmony v1.6 has no published score. Instrument refinement remains experimental, with datasets shown separately.
| Metric rank | Model | Test setting | DrumKit F1 (%)↑ | AllPercussion F1 (%)↑ |
|---|---|---|---|---|
| 1 | drums_v1_5Public weights · Inference code | Audio-aligned · 50 ms · canonical GM | 68.9 | 40.44 |
| 2 | drumsPublic weights · Inference code | Audio-aligned · 50 ms · canonical GM | 61.57 | 18.79 |
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 -.
Tsumugi Other and Vocal Harmony report COnP, not COnF1. Published results cover v1.5; default Vocal Harmony v1.6 has no published score. Instrument refinement remains experimental, with datasets shown separately.
| Metric rank | Model | Test setting | ACC (%)↑ |
|---|---|---|---|
| - | Instrument refinement modelPublic weights · Inference code | Held-out RWC-I | 74.5 |