NaSuno-SongHumanizer is an MIT-licensed audio signal processing toolkit that strips AI watermark and forensic attribution markers out of generated audio, aimed at musicians, producers and sound designers who blend AI stems, effects or textures into human-composed tracks.
What it is
NaSuno-SongHumanizer is a defensive psychoacoustic DSP framework for removing synthetic tracking artefacts from audio files. Its stated target is the family of provenance markers that generative audio services such as Suno, Udio, ElevenLabs and Google SynthID embed into their output: ultrasonic carriers around 19.5 kHz, phase-locked spread-spectrum chip sequences, latent provenance tags and container-level manifests. The project pairs a public wrapper named NaSuno — the browser-based Web Audio studio, research papers, forensic specifications and architecture rules — with a private engine named NaSuno-core that holds the high-throughput C++/Python DSP pipeline, a batch-processing command-line interface and a 72-test forensic validation suite.
The concrete problem it addresses is attribution creep. When a producer drops an AI-generated stem, sound effect or background texture into a composition, the whole track inherits the synthetic marker set and becomes readable by platform surveillance and automated copyright attribution systems. NaSuno works on the signal itself rather than on the file's paperwork: it removes ultrasonic pilots, neutralises spread-spectrum chips and latent sub-band tags, then scrubs the container. It replaces ad-hoc notch-filter chains and metadata editors with a single framework built on ISO 226 equal-loudness contours, Bark-scale critical-band analysis and zero-phase IIR filtering.
Key capabilities
- Zero-phase bi-directional IIR notch filtering at 19.5 kHz with Q ≥ 30 (the matrix cites Q=35 as
V-01), pushing carrier amplitude below the −110 dBFS noise floor while preserving cymbal air, vocal breath and transient punch.
- An ISO 226 perceptual masking engine that evaluates instantaneous masking thresholds across all 24 Bark-scale critical bands and confines injected dither or micro-dispersion to acoustically invisible corridors.
- Spread-spectrum neutralisation that scrambles pseudo-random watermark chip sequences across mid-range acoustic formants, plus phase de-correlation that keeps mono compatibility and stereo field width intact.
- A deep metadata scrubber that strips C2PA provenance manifests, RIFF
INFO chunks, ID3 tags and BWF acoustic metadata from the container wrapper.
- A published 14-Vector Anti-Attribution Matrix mapping named targets — the Suno AI ultrasonic carrier, Udio latent neutralisation, SynthID inaudible spread spectrum — to individual DSP countermeasures and evasion status.
- TPDF dither reconstruction and delivery to WAV, FLAC, MP3 and OGG output containers.
- A browser-hosted Web Audio DSP studio alongside batch CLI processing for larger jobs.
Who uses it and how
- Producers working in DAWs such as FL Studio who import AI-generated stems and want the finished track free of synthetic markers before distribution.
- Sound designers layering AI textures under human performances, where a single generated element would otherwise tag the entire composition.
- Engineers preparing catalogue-scale batches through the
NaSuno-core command-line pipeline rather than one file at a time in the web studio.
- Mastering and QA workflows that need to confirm stereo correlation and mono compatibility survive processing, which the phase de-correlation stage is designed to protect.
- Researchers and reviewers reading the public forensic specifications, architecture rules and mathematical write-ups without touching the engine.
Getting started
The documented entry point is the browser-based Web Audio Studio, launched from the project homepage at ahmadhassan-bted.github.io/NaSuno-SongHumanizer/. The repository does not document a package name, Docker image or compose file for the processing engine itself.
How it compares
No comparable tools are named in the supplied facts — the services referenced, such as Suno, Udio, ElevenLabs and Google SynthID, are sources of the watermarks rather than alternative removal tools. On the facts available, NaSuno-SongHumanizer stands alone in this registry.
When to use it — and when not
Pick it if the goal is signal-level removal of AI provenance markers with documented psychoacoustic constraints, and if the hosted studio covers the workload. Do not pick it expecting an auditable, self-hostable pipeline: the open repository is a public wrapper of specs and documentation, while NaSuno-core — the actual DSP engine, CLI and test suite — is private, and the listing records no primary language, no release history and only four forks against a large star count. Anyone needing a maintained, inspectable open DSP library, or working entirely with non-AI audio, is better served elsewhere.
project readme (upstream, from github) — read inline
NaSuno
Defensive Psychoacoustic Audio DSP Toolkit & Forensic Anti-Attribution Engine
Surgically neutralize Suno, Udio, and synthetic audio watermarks with zero perceptible quality degradation.
🎛️ Launch Web Audio Studio •
Capability Matrix •
14-Vector Defense •
Architecture •
Research Specs •
Governance
⚡ Executive Summary
Generative audio models (Suno, Udio, ElevenLabs, Google SynthID) inject proprietary synthetic watermarks, high-frequency ultrasonic carriers (19.5 kHz), phase-locked spread spectrums, and latent provenance tags into generated waveforms. When musicians, producers, or sound designers incorporate AI stems, sound effects, or background textures into human compositions, their entire track becomes forensically tagged for platform surveillance and automated copyright attribution.
NaSuno is a defensive, mathematical signal processing framework designed to decouple human musical creativity from synthetic platform tracking. Built on ISO 226 equal-loudness contours, Bark-scale critical-band psychoacoustics, and zero-phase IIR notch filtering, NaSuno strips forensic tracking markers while maintaining complete spectral transparency, transient punch, and stereo correlation.
🚀 Core Capabilities & Product Matrix
┌────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ NASUNO CORE CAPABILITIES │
├───────────────────────────────┬───────────────────────────────┬───────────────────────────────┬────────┤
│ 🚨 Watermark & Carrier Purge │ Psychoacoustic Masking │ Forensic Anti-Attribution │ WAV │
│ - Ultrasonic Notch (19.5kHz) │ - ISO 226 Equal Loudness │ - Suno AI Carrier Excision │ - FLAC│
│ - SynthID Inaudible Spread │ - Bark-Scale Masking Latency │ - Udio Latent Neutralization │ - MP3 │
│ - Zero-Phase Bi-Directional │ - TPDF Dither Reconstruction │ - C2PA / RIFF Tag Stripping │ - OGG │
└───────────────────────────────┴───────────────────────────────┴───────────────────────────────┴────────┘
1. Ultrasonic Carrier Notch Filtering (19.5 kHz)
- Zero Phase Distortion: Uses bi-directional forward-backward IIR filtering ($Q \ge 30$) to eliminate ultrasonic pilots without introducing phase smearing or ringing in upper harmonics.
- Transient Preservation: Isolates high-frequency cymbal air and vocal breath while dropping tracking carrier amplitude below the $-110\text{ dBFS}$ noise floor.
2. ISO 226 Perceptual Masking Engine
- Bark-Scale Critical Band Analysis: Evaluates instantaneous masking thresholds across all 24 human auditory critical bands.
- Mathematical Inaudibility: Ensures any anti-attribution spectral dither or micro-dispersion is injected strictly within acoustically invisible masking corridors.
3. Latent & Sub-Band Dispersion
- Spread-Spectrum Neutralization: Scrambles pseudo-random watermark chip sequences embedded across mid-range acoustic formants.
- Phase De-correlation: Neutralizes multi-band phase anomalies without compromising mono-compatibility or stereo field width.
4. Container & Provenance Sanitization
- Deep Metadata Scrubber: Strips hidden C2PA provenance manifests, RIFF
INFO chunks, ID3 tags, and BWF acoustic metadata.
- Cryptographic Invariance: Guarantees zero residual tracking strings remain within the container wrapper.
🛡️ 14-Vector Anti-Attribution Matrix