Vital Audio JEPA V6 One-Shots

V6 continues the V5 encoder on controlled one-shot articulations rendered by Vital 1.6.4. It is intended to improve transient, envelope, tail, spectral, and stereo representations before retraining the Vital preset decoder.

Input and corpus

  • 22,050 Hz stereo PCM16, exactly 6.0 seconds
  • one-second note gate plus five-second decay/release region
  • no amplitude normalization
  • 80,146 accepted one-shots from 7,394 parent presets
  • 17 controlled envelope/pitch candidates per parent before filtering
  • parent-disjoint train, validation, and test splits

The dataset is stored under oneshots_v0_22050_stereo_6s/ in duckking032/vital.

Training

  • initialized from duckking032/vital-jepa-v5
  • all V5 weights loaded with no missing or unexpected keys
  • 36,519,146 parameters total
  • 76,260 train / 1,974 validation / 1,912 test clips
  • 10 epochs / 11,910 steps
  • selected checkpoint: step 10,000

The pretrained encoder used a lower learning rate than the auxiliary heads. Training optimized JEPA prediction, acoustic descriptors, spatial descriptors, coarse category supervision, and category contrastive structure.

Untouched one-shot test results

Metric V5 baseline V6
Masked JEPA loss 0.02317 0.01351
Acoustic descriptor correlation 0.5949 0.7618
Acoustic descriptor MAE 0.0646 0.0418
Coarse category accuracy 18.20% 48.01%
Coarse retrieval top-1 31.07% 31.64%
Coarse retrieval top-5 56.28% 57.37%
Coarse retrieval MRR@10 0.4218 0.4284

The largest gains are in tail energy, tail-to-body ratio, attack time, spectral descriptors, and channel imbalance. Modulation-depth correlation remains effectively unresolved on this one-shot corpus.

Limitations

This continuation corpus contains articulation variants but no unchanged performance reference clips or exact single-control pairs. V6 therefore improves one-shot acoustic representation but does not replace V5's exact-control evaluation. The downstream decoder should retain V5-compatible control heads and be retrained on a mixture of exact-control and one-shot data.

checkpoints/jepa_v6_oneshots/selected.pt is the optimizer-free selected checkpoint. The checkpoint comparison and both V5/V6 one-shot evaluation reports are included.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train duckking032/vital-jepa-v6-oneshots