Negative Sound Wave — AI-artifact suppression via subtractive common-mode rejection

[ophiocus]

Feature Request — TinyBooth · Negative Sound Wave (AI-artifact suppression)

Product: TinyBooth (Sound Studio) · Type: correction item / DSP feature · Status: proposed
Origin: operator seat (D:\tecnocratica) → routes to dev via ophiocus/tecnocratica-ops issue.
One-liner: derive an anti-artifact signal from a track — the "negative wave" — that, subtracted,
suppresses the generative (Suno/AI) residue and restores each track's separation and clarity.

---

1 · Problem (the "shrink-wrap" / underwater effect)

AI music (Suno et al.) is generated from noise by a diffusion/transformer decoder. The decoder never
perfectly reconstructs phase and fine time–frequency structure, so it leaves a correlated residue baked
into every output — an allegorical "shrink-wrap." Symptoms the operator hears:

• "Underwatery" — phase smearing / warble from imperfect vocoder phase reconstruction.
• Loss of "singularity" — when a song is split into stems, each stem carries shared generative haze
  and bleed from the others, so the tracks don't feel distinct; they sound fused, flat, smeared together.

Key realization that makes this solvable: these artifacts are statistical outliers, not music.
They are (a) common across stems (shared decoder residue), (b) phase-incoherent with real musical
partials, and (c) spectrally diffuse (a haze floor, not tonal content). Anything that satisfies all
three is almost certainly artifact — and can be estimated and subtracted without touching the music.

2 · Concept — the Negative Sound Wave

Estimate the artifact content  of a track, invert it, and mix it back so it destructively interferes
with the residue (noise-cancellation logic, but targeted at AI artifacts rather than ambient noise).
The estimated artifact, phase-inverted, is the "negative wave" — and TinyBooth should expose it as a
soloable, audible signal so the operator can hear what is being removed and confirm it's residue,
not music, before committing. A single Depth knob controls how much is subtracted.

3 · How to find the outliers — the sound logic & math

Work in the STFT domain. Let X(t,f) be a track's complex spectrogram (magnitude |X|, phase ∠X).
Estimate an artifact magnitude profile Â(t,f) from as many of these estimators as apply, then combine:

(A) Cross-stem common-mode rejection — the flagship (needs stems).
The shared decoder residue is the part of each stem predictable from the others — i.e. the common mode.
With K stems X_k(t,f), estimate the common component per bin (robust: C = median_k |X_k|, or the
dominant cross-stem singular vector via a per-bin rank-1 SVD across the stem matrix), gated by cross-stem
phase coherence (only bins where stems agree count as shared). Subtract C from each stem. This is a
differential amplifier rejecting common-mode noise — it decorrelates the stems, which is literally the
"restore singularity" the operator wants: less shared haze ⇒ more separation between tracks.

(B) Phase-incoherence mask (single-track). Real partials evolve with coherent instantaneous frequency;
vocoder warble does not. Compute a per-bin phase-coherence / group-delay stability metric; low-coherence
bins are artifact → include in Â. Targets the "underwatery" smear directly.

(C) Minimum-statistics noise floor (single-track). Track the spectral floor over time (Martin's minimum
statistics). The persistent floor beneath the music is the diffuse generative haze.

(D) HPSS residual (single-track). Median-filter the spectrogram along time (→ harmonic) and along
frequency (→ percussive); the leftover third component (neither stable-tonal nor transient) is largely
the AI residual. Feed it into Â.

Combine → subtract → reconstruct:

Â(t,f)   = weighted blend of the estimators above (coherence-gated)
|Ŝ|      = max( |X| − α·Â ,  β·|X| )         # α = Depth (over-subtraction); β = spectral floor
Ŝ        = |Ŝ| · e^{j·∠X}   →  iSTFT          # cleaned track (keep original phase)
negWave  = iSTFT( α·Â · e^{j·∠X} )            # the audible "negative wave": Ŝ = X − negWave

Anti-"musical-noise" safeguard (mandatory). Naïve spectral subtraction introduces warbly "musical
noise" — worse than the disease. Use a decision-directed / log-MMSE (Ephraim–Malah) gain instead of raw
subtraction, plus temporal smoothing and the spectral floor β, so removal is smooth and gated by an a-priori
SNR estimate rather than punching holes. This is the difference between a real feature and a toy.

4 · UX / controls

• Negative Wave: solo / mute / gain — hear the residue being removed; A/B the cleaned track.
• Depth (α) — global amount. Floor (β) — protect against over-scrubbing.
• Estimator mix — advanced: weight common-mode vs phase vs floor vs HPSS (sensible defaults hidden).
• Presets — "Suno stem," "Suno full mix," "gentle de-haze," "aggressive separate."
• Null-test button — invert cleaned against original to audition exactly what changed.

5 · Phasing

• MVP: single-track de-haze — estimators (B)+(C)+(D), log-MMSE subtraction, soloable negative wave,
  Depth + Floor knobs, one "Suno" preset. Ships the core value on any single track.
• Full: multi-stem common-mode rejection (A) with coherence gating (the "restore singularity"
  headline), estimator-mix panel, per-source presets, batch across a session's stems.

6 · Non-goals / risks

• Not a general denoiser, mastering chain, or upsampler; it removes residue, it does not hallucinate
  back lost detail.
• Risk: stripping intended airy/ambient/reverb content — mitigated by the audible negative wave +
  conservative defaults + Floor. Risk: musical noise — mitigated by log-MMSE (see §3). Risk: CPU on
  long stems — offline/render-time acceptable; realtime is a stretch goal.

7 · Acceptance criteria

• Negative wave is soloable and, at tuned defaults, contains no clearly audible musical content on a
  reference set (it's residue).
• Measurable separation gain: inter-stem correlation drops after common-mode rejection; noise-floor
  spectral flatness improves; A/B null test isolates only residue.
• No audible musical noise introduced on the reference set at default Depth.
• Cleaned tracks judged clearer / less "underwatery" in blind A/B by the operator.

---

🇨🇴 Resumen

Onda Sonora Negativa para TinyBooth: deriva de una pista una señal anti-artefacto (la "onda negativa")
que, al restarse, suprime el residuo generativo de la IA (Suno) — ese "shrink-wrap" que deja las pistas
"submarinas" y fundidas entre sí (sin singularidad). La clave: los artefactos son outliers
estadísticos — (a) comunes entre stems, (b) incoherentes en fase con los parciales musicales, y
(c) difusos (un piso de ruido, no tonos). Lo que cumple los tres es artefacto y se puede estimar y restar
sin tocar la música. Método (dominio STFT): el estimador estrella es el rechazo de modo común entre
stems (un amplificador diferencial: quita lo compartido ⇒ decorrelaciona las pistas ⇒ recupera la
singularidad), apoyado por máscara de incoherencia de fase, piso de ruido de mínima estadística y
residuo HPSS. Se resta con ganancia log-MMSE (Ephraim–Malah) — no resta cruda — para evitar el "ruido
musical" warbly. UX: la onda negativa es audible/soloable (oír qué se quita antes de aplicar) + perilla
de Profundidad y Piso + presets Suno. MVP: de-haze de pista única; Full: rechazo de modo común
multi-stem. No es un denoiser general ni reconstruye detalle perdido.