NaSuno is a psychoacoustic DSP toolkit that detects and neutralizes ultrasonic carriers, SynthID spread-spectrum marks, and C2PA provenance metadata in Suno and Udio output, aimed at authorized forensic and production research.
| Tool | AhmadHassan-BTed/NaSuno-SongHumanizer — psychoacoustic anti-attribution and AI watermark analysis toolkit with a browser-based DSP studio |
| Category | Audio forensics / signal processing / steganalysis |
| Primary Use | Analyzing and neutralizing embedded watermarks (19.5 kHz carriers, SynthID chips, C2PA manifests) in AI-generated audio within labs and research environments |
| Safe Use | For authorized forensic researchers, audio engineers studying watermark robustness, and defensive teams auditing provenance signals on content they own or are licensed to process |
| Telemetry Note | The web studio is documented as 100% client-side (Web Audio API, in-memory, nothing uploaded), so defenders should instead watch for watermark absence: platform-side detection observes when provenance carriers disappear from redistributed audio |
NaSuno, hosted at AhmadHassan-BTed/NaSuno-SongHumanizer under an MIT license, is one of the more technically articulate entries in the emerging category of AI-audio provenance tooling. Its premise is straightforward: generative audio platforms such as Suno, Udio, ElevenLabs, and Google SynthID embed forensic watermarks into their output — ultrasonic carriers around 19.5 kHz, phase-locked spread-spectrum sequences, and latent provenance tags — and those markers survive into derivative works. NaSuno is a defensive, mathematical DSP framework designed to study and neutralize those markers, built on ISO 226 equal-loudness contours, Bark-scale critical-band psychoacoustics, and zero-phase IIR notch filtering. With nearly six thousand stars and a Trendshift badge, it has clearly struck a nerve among producers and forensic researchers alike.
What the README actually documents is less a single binary and more a layered system. The public repository is a wrapper: it hosts a browser-based Web Audio DSP studio, mathematical research papers in docs/research/, an architecture specification in docs/architecture.md, and a 14-vector forensic taxonomy of watermark mechanisms. The high-throughput processing pipeline — a C++/Python DSP core, a batch CLI, and a 72-test verification suite — is described as a private component (NaSuno-core) referenced but not shipped in this repo. That matters for evaluation: what you can clone today is the specification, the research corpus, and the interactive studio, not the full batch engine.
The technical core rests on a few well-chosen primitives. The first is an ultrasonic carrier notch at 19.5 kHz implemented with bi-directional forward-backward IIR filtering at Q ≥ 30, which cancels the pilot tone without the phase smearing a single-pass filter would introduce. The README claims carrier amplitude drops below the -110 dBFS noise floor while preserving cymbal air and vocal breath — a plausible claim given that a zero-phase filtfilt-style approach avoids the pre-ringing artifacts that would betray the processing. The second primitive is the ISO 226 perceptual masking engine, which computes instantaneous masking thresholds across all 24 human auditory critical bands so that any anti-attribution dither injected by the tool stays inside acoustically invisible corridors.
The 14-vector matrix is where the project shows its forensic ambition. Beyond the headline Suno and Udio carriers, it catalogues SynthID spread-spectrum chip sequences neutralized through Bark-band de-correlation, neural vocoder comb peaks from HiFi-GAN and EnCodec smoothed by an adaptive spectral filter, infrasonic 5–18 Hz pilots removed with a 24 dB/Oct Butterworth high-pass at 20 Hz, and deterministic quantization footprints addressed through TPDF psychoacoustic dithering. Several vectors are less about watermark removal and more about de-fingerprinting the generative signature itself: rigid micro-timing lattices get Gaussian jitter, flat dynamic-range curves get transient expansion, and unnatural spectral-flux regularity receives entropic modulation. Read as a taxonomy document alone, the matrix is a genuinely useful survey of how synthetic audio betrays its origin.
Container-level sanitization is handled by a separate metadata scrubber that strips C2PA provenance manifests, RIFF INFO chunks, ID3 tags, and BWF broadcast-wave metadata, with zero-byte re-alignment producing what the README calls a bit-perfect header rewrite. This is the layer defenders should care about most, because C2PA is the industry's emerging content-provenance standard; a tool that excises those manifests cleanly is directly relevant to anyone researching how robust provenance chains actually are once media leaves controlled pipelines. The interplay between the spectral vectors (V-01 through V-03) and the container vectors (V-04, V-11) illustrates the real defense problem: provenance must survive both signal degradation and file restructuring, and this project methodically probes both halves.
The web studio is the part most readers will actually touch. Hosted via GitHub Pages from docs/index.html, it processes audio entirely client-side through the Web Audio API — no uploads, no logging — inside an FL Studio-inspired skeuomorphic rack with rotary knobs, LED VU meters, a live 2048-point FFT analyzer, and a stereo phase oscilloscope. A single humanization dial sweeps from 0% raw signal to 100% fully processed master, which makes it a usable demonstration environment for hearing (or, by design, not hearing) what each processing stage does. For a defensive researcher, that immediacy is valuable: you can feed in a watermarked sample you generated yourself and observe the spectral deltas in real time without writing any code.
Architecturally, the README commits to hexagonal architecture — ports, adapters, explicit audio-buffer contracts, and latency budgets — with the public repo acting as specification and interface layer while the private NaSuno-core implements the engine. CI workflows (ci.yml, pages.yml) verify documentation and the JSON registry and deploy the studio automatically. The docs/ tree is the substantive payload: forensic-taxonomy.md, psychoacoustics-reference.md, master-evasion-plan.md, api-reference.md, and a roadmap.md targeting VST3/AU plugin builds and hardware DSP ports. The roadmap signals the author intends this to land inside conventional DAW workflows as an audio plugin rather than remain a research curiosity.
Now the honest caveat, because it deserves one. Despite the defensive framing, the tool's practical gravity points toward evading platform attribution and automated copyright systems, and stripping C2PA manifests from content you did not generate or are not licensed to process sits in legally and ethically contested territory — it can defeat provenance systems designed to protect creators and to detect synthetic media. The legitimate authorized contexts are real: robustness testing of watermark schemes you operate, forensic analysis of synthetic-media detection, and processing your own AI-assisted compositions where platform terms allow it. But a senior operator should treat NaSuno the way one treats nmap or a C2 framework — a capable instrument whose value depends entirely on the authorization behind the keyboard. Anyone using it to launder copyrighted output or defeat provenance protections is operating outside those bounds.
For defenders, the tool is more useful as a map of the attack surface than as a threat itself. Every vector in the matrix tells a detection team where provenance signals are fragile: ultrasonic carriers die to resampling and notching, spread-spectrum chips to phase de-correlation, metadata to container rewrites. Any watermarking or provenance deployment that relies on a single fragile channel will not survive contact with tooling like this, which argues for layered, redundant provenance combining in-band signal marks, C2PA manifests, and distribution-side telemetry. NaSuno's documentation is, in effect, a free red-team report on AI audio watermarking — and it is worth reading on those grounds alone.
AhmadHassan-BTed/NaSuno-SongHumanizer.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.