Vital JEPA V8
Stereo 22.05 kHz, six-second audio encoder initialized from V7 and trained for 20,000 steps on native Vital interventions, synthetic one-shot articulations, and a licensed acoustic pilot corpus.
V8 adds paired acoustic views, acoustic instance contrastive learning, supervised MIDI-pitch regression, descriptor-distance metric alignment, channel swaps, polarity flips, small gain changes, and small time shifts. Audio is loaded as stereo without normalization.
The selected checkpoint is step 20,000. It was selected for its overall downstream profile despite step 16,000 having a marginally lower masked validation loss.
V7 to V8 acoustic evaluation
The fixed evaluation set contains 150 acoustic sounds held out by source group.
| Metric | V7 | V8 |
|---|---|---|
| Masked prediction loss | 0.06353 | 0.04550 |
| Fine-family top-1 | 84.00% | 82.67% |
| Fine-family top-5 | 93.33% | 96.00% |
| Descriptor MAE | 0.07059 | 0.03518 |
| Descriptor correlation | 0.7392 | 0.8160 |
| Normalized pitch MAE | 0.09298 | 0.04664 |
V8 is substantially stronger on acoustic pitch and descriptors. It is not an across-the-board V7 replacement: on the untouched 5,599-item synthetic Vital test split, fine top-1 changed from 36.97% to 36.63%, top-5 from 58.81% to 57.12%, descriptor correlation from 0.8703 to 0.8596, and masked loss from 0.04980 to 0.05850.
Use V8 for acoustic-input decoder experiments, and retain V7 as the synthetic
Vital-domain baseline. See evaluation/ for full machine-readable reports.
Selected checkpoint SHA-256:
1fdc86f16eea9f32557cb2e71cb0b463249e8e247c00446f6bf77a91095cd70f.