Vital Audio JEPA V1
Masked audio JEPA encoder trained on native Vital 1.6.4 renders from
duckking032/vital.
Input
- 22,050 Hz mono
- 6.0 seconds
- 128-bin log-mel spectrogram
- 1024-sample FFT, 256-sample hop
- 16x16 spectrogram patches (8x32 grid, 256 tokens)
The source renders were produced by loading unchanged .vital files in Vital's
standalone headless renderer. Training audio was averaged to mono, anti-aliased
downsampled from 44.1 kHz, and trimmed to six seconds. Near-silent renders,
severely clipped renders, and exact duplicate audio were excluded.
Objective
The online encoder receives 40% context patches. A predictor reconstructs 40% target patches in the latent space of an exponential-moving-average target encoder.
This model is intended as a retrieval and initialization baseline for editable Vital sound-design workflows, not exact preset reconstruction.
See eval_runs/jepa_v1/results/eval_report.json for measured masked loss and
same-category retrieval metrics.
Evaluation
The selected final checkpoint was evaluated deterministically on all 7,394 filtered clips:
- masked validation loss: 0.10173
- top-1 same-type retrieval: 39.57%
- top-5 same-type retrieval: 71.42%
- mean reciprocal same-type rank at 10: 0.5301
For context, V0 reported 24.4% top-1 and 53.8% top-5 retrieval, but it was trained on the subsequently rejected approximation renders, so that comparison is directional rather than controlled.
The final checkpoint was selected for retrieval quality. The step-10,000 checkpoint had a marginally lower masked validation loss (0.10133) but slightly weaker retrieval.