Audio Forensics & Restoration
AI-generated music often suffers artifacts from lossy training data — 16 kHz ceilings, phase incoherence, spectral smearing. Six forensic tells identify it, and a diagnostic plus restoration path was proposed to repair it.
Artifact Index
| Tell | Symptom |
|---|---|
| HF shelf | Hard 16/22 kHz cutoff |
| Transients | Smeared drum attacks |
| Stereo | Collapsed or pseudo-width |
| Phase | Incoherent left/right correlation |
| Vocals | Robotic formants |
| Noise | Pumping artifacts |
Rehab Pipeline
Historical proposal
This three-step path from AI generation to release-grade fidelity is design documentation, not current executable MixBox behavior.
- Diagnose — detect specific issues, e.g. with
mastering/scripts/diagnose.pyin the mastering pipeline. - Neural super-resolution — break the 16 kHz ceiling; theorized to run before restoration.
- Unified restoration — fix dynamic, spatial, and EQ degradations.
SR Tools
| Tool | Niche |
|---|---|
| AudioSR | Most versatile super-resolution; cloud/GPU preferred, or CPU batch |
| LavaSR | Local CPU-only processing, ~50x realtime |
| NovaSR | Ultra-fast fallback |
Ownership & Legal Strategy
Historical
Raw AI output sits in a legal grey zone; establishing the human master in a DAW was treated as a requirement for copyright defense and streaming-platform acceptance.