A DSP rack that claims to scrub AI fingerprints off your music

What it does NaSuno targets the forensic markers generative audio platforms allegedly embed in their output: ultrasonic carriers around 19.5 kHz, spread-spectrum chip sequences, metadata manifests, and neural-vocoder artifacts. It counters these with a 14-vector matrix of DSP techniques — zero-phase IIR notching, Bark-scale perceptual masking, TPDF dithering, micro-timing jitter, and metadata scrubbing. The public repo ships a fully client-side Web Audio studio styled like an FL Studio rack, with a live FFT analyzer and a “humanization” dial from raw AI signal to humanized master.

The interesting bit The repo is deliberately split: this public half holds the browser studio and research specs, while the actual C++/Python DSP pipeline lives in a private “NaSuno-core” engine. So the most load-bearing code — the thing doing the watermark removal — is not in the repository you’re staring at.

Key highlights

• Runs entirely in-browser via Web Audio API; no uploads, no logging, no installation
• 14-vector countermeasure matrix covering carriers, phase artifacts, comb peaks, and container metadata (C2PA, RIFF, ID3)
• ISO 226 equal-loudness and Bark-band psychoacoustics used to keep interventions inaudible
• Research docs include a forensic watermark taxonomy and an evasion strategy paper
• MIT-licensed, with a roadmap pointing at VST3/AU plugin targets

Caveats

• The core engine is private, so the claims can’t be verified from this repo alone
• The 14-vector table asserts “Neutralized” status for every vector, but no independent validation or benchmark results are shown
• The exact watermark mechanisms of Suno, Udio, and SynthID are not public, so the taxonomy is partly speculative

Verdict Interesting for producers curious about AI-audio provenance and psychoacoustic DSP, and the browser studio is a fun toy. But anyone needing provable watermark removal should note the engine doing the proving is behind a curtain.