Evaluation Results
Classification (5-fold CV)
| Task | Model | Acc |
|---|---|---|
| Pianist8 | Aria | 0.861+/-0.043 |
| Pianist8 | Moonbeam | 0.806+/-0.045 |
| Pianist8 | MusicBERT | 0.786+/-0.021 |
| Pianist8 | CLaMP3 | 0.808+/-0.033 |
| Pianist8 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.831+/-0.032 |
| Pianist8 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.868+/-0.052 |
| Pianist8 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.842+/-0.038 |
Ranking (5-fold CV)
| Task | Model | PairAcc |
|---|---|---|
| chopin_comp | Aria | 0.475+/-0.081 |
| chopin_comp | Moonbeam | 0.537+/-0.029 |
| chopin_comp | MusicBERT | 0.525+/-0.058 |
| chopin_comp | CLaMP3 | 0.535+/-0.085 |
| chopin_comp | MAJEPPA ($\mathcal{L}_{piece}$) | 0.537+/-0.081 |
| chopin_comp | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.587+/-0.087 |
| chopin_comp | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.597+/-0.094 |
Technique Classification (5-fold CV)
| Task | Eval | Model | Acc |
|---|---|---|---|
| technique_cls | single-label | Aria | 0.662+/-0.114 |
| technique_cls | single-label | Moonbeam | 0.779+/-0.048 |
| technique_cls | single-label | MusicBERT | 0.671+/-0.046 |
| technique_cls | single-label | CLaMP3 | 0.703+/-0.027 |
| technique_cls | single-label | MAJEPPA ($\mathcal{L}_{piece}$) | 0.763+/-0.059 |
| technique_cls | single-label | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.703+/-0.095 |
| technique_cls | single-label | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.790+/-0.065 |
Performance Assessment (5-fold CV)
| Task | n | Model | MAE |
|---|---|---|---|
| pisa | 59 | Aria | 0.122+/-0.031 |
| pisa | 59 | Moonbeam | 0.151+/-0.035 |
| pisa | 59 | MusicBERT | 0.190+/-0.026 |
| pisa | 59 | CLaMP3 | 0.217+/-0.036 |
| pisa | 59 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.125+/-0.083 |
| pisa | 59 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.108+/-0.090 |
| pisa | 59 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.090+/-0.034 |
| neuropiano | 104 | Aria | 0.214+/-0.019 |
| neuropiano | 104 | Moonbeam | 0.192+/-0.014 |
| neuropiano | 104 | MusicBERT | 0.205+/-0.019 |
| neuropiano | 104 | CLaMP3 | 0.203+/-0.017 |
| neuropiano | 104 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.180+/-0.015 |
| neuropiano | 104 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.181+/-0.022 |
| neuropiano | 104 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.155+/-0.018 |
| ycuppe | 606 | Aria | 0.068+/-0.005 |
| ycuppe | 606 | Moonbeam | 0.061+/-0.004 |
| ycuppe | 606 | MusicBERT | 0.065+/-0.004 |
| ycuppe | 606 | CLaMP3 | 0.089+/-0.004 |
| ycuppe | 606 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.064+/-0.004 |
| ycuppe | 606 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.062+/-0.005 |
| ycuppe | 606 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.059+/-0.002 |
UMP -- Unreferenced Mistake Detection
| Dataset | n | Model | F1 (mean±std) |
|---|---|---|---|
| EN-augmented | 34 | Aria | 0.174+/-0.025 |
| EN-augmented | 34 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.151+/-0.021 |
| EN-augmented | 34 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.141+/-0.015 |
| EN-augmented | 34 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.179+/-0.012 |
| Burgmuller | 25 | Aria | 0.313+/-0.016 |
| Burgmuller | 25 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.308+/-0.012 |
| Burgmuller | 25 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.321+/-0.012 |
| Burgmuller | 25 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.333+/-0.015 |
| ScoreInformed | 7 | Aria | 0.300+/-0.020 |
| ScoreInformed | 7 | MAJEPPA ($\mathcal{L}_{piece}$) | 0.313+/-0.015 |
| ScoreInformed | 7 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}$) | 0.341+/-0.004 |
| ScoreInformed | 7 | MAJEPPA ($\mathcal{L}_{piece}+\mathcal{L}_{cond}+\mathcal{L}_{self}$) | 0.316+/-0.037 |
UMP Mistake Prediction Examples
Comparing the Aria-medium baseline with MAJEPPA representations on UMP mistake predictions.
Example 1
Example 2
Example 3
Example 4
