🥣Instant Soup/ Poison Pill
all-local · nothing uploadeddetecting backend…
🌶️ Anti-AI Seasoning for Musicians

Poison the Pot,
Protect the Recipe.

Drop a track. We stir in a hidden layer of anti-AI noise, shaped so your ears don't notice — but generative music models that try to learn from your song get a mouthful of bad broth. Everything simmers right here in your browser. Nothing gets uploaded.

the recipe, in plain english

  1. 1.
    Pick a strength. Stronger = harder for AI to learn from, but slower and a slight risk of a faint hiss.
  2. 2.
    We fold in carefully-crafted "confusion noise" and keep it under a hearing curve so it hides inside your own music.
  3. 3.
    You get back a WAV that sounds the same to fans, but is much less useful as training data for music-generating AI.

the kitchen · engine

surrogate model
Note: the surrogate is a real convolutional network — gradients truly propagate through STFT → mel → conv stack → embedding + logits into the waveform. Weights are seeded-random, matching the shapes of YAMNet / PANNs-CNN10. Loading pretrained weights is on the roadmap.

01 · ingredients

or drop a file anywhere in this panel. WAV / FLAC / MP3 / M4A accepted; output is 16-bit WAV.

02 · pick your spice level

A good middle ground for most releases. Strong enough to disrupt AI training, still masked by your track so listeners won't notice.
01
Feature-destruction loss
Perturbations push log-mel band energies away from the clean signal — the same target used by Harmony-Cloak-style attacks.
02
Psychoacoustic mask
Every step projects the added noise below a per-frame ATH + Bark-spread masking threshold so it stays under audibility.
03
Neural surrogate CNN
v2 runs true PGD with gradients through a convolutional audio tagger (YAMNet-scale or PANNs-scale) on WebGPU. Combined loss destroys the model's embedding and misdirects its classifier head.