Phase I · Institutional Research Tier
Wavelet scattering, shipped as a wheel.
Drop-in C++/CUDA primitives for any audio, vibration, or biosignal pipeline that needs translation-invariant, frequency-aware fingerprints. AGPL-3.0 + Commercial.
WSTConfig
Depth
16,512 scattering paths
at depth 2
at depth 2
Time domain: λ=0
Frequency domain: bank
Filter λ=0 · ξ=π·2^(-0/16) · σ=0.800 samples · peak=0.051 (analytic Morlet, normalised)
What you get
- →omni-wst-core PyPI wheel: WSTEngine<HopperTag, J, Q> + analytic Morlet bank + Radix-2 FFT CPU fallback.
- →JTFS (joint time-frequency scattering) on dual CUDA streams.
- →15+ GB/s DMA via cudaHostRegister on pinned Plasma pages.
What it solves
- Spectrogram-and-CNN pipelines burn enormous compute to learn what scattering provides analytically.
- WST gives you orderable, comparable fingerprints with formal stability guarantees, in milliseconds.
Pricing
AGPL-3.0: free for research & open source
Enterprise use cases
Audio QA / royalty enforcement
Detect derivative tracks in a 100M-asset catalogue using deterministic fingerprints.
Bioacoustic monitoring
Species and event classification on field recordings without retraining for each habitat.
Industrial vibration analytics
Spot the fault signature in rotating machinery before a model would even converge.
Quickstart
from omni_wst_core import fingerprint, WSTConfig fp = fingerprint(signal, WSTConfig(J=8, Q=16, depth=2, jtfs=False))