Architecture
How it works
ArtifactNet reframes AI music detection as forensic residual physics. Instead of learning generator-specific stylistic fingerprints, it isolates the physical residue that neural audio codecs imprint on generated audio. A bounded-mask UNet extracts those codec residuals from the magnitude spectrogram, HPSS decomposes them into multi-channel forensic features, and a compact CNN classifies real vs. AI-generated.
STFT
Magnitude
spectrogram
ArtifactUNet
Bounded-mask UNet
codec residual extractor
HPSS decomp.
Median-filter HPSS →
7-channel forensic features
CNN classifier
Compact CNN
residual structure scoring
Verdict
Real vs. AI
+ calibrated confidence
The mechanism: commercial AI music generators (Suno, Udio, Stable Audio, MusicGen, Riffusion) all rely on neural audio codecs that use Residual Vector Quantization (RVQ), mapping continuous audio onto discrete codebook vectors. When a separation model trained only on human music meets AI-generated audio, the reconstruction residuals have measurably different structure — a phenomenon the paper terms forensic residual amplification. ArtifactNet detects that structural difference, so it generalizes to generators never seen in training. Read the paper (arXiv:2604.16254) →
RVQ residue
Residual Vector Quantization leaves structured codebook-quantization residue across the spectrum. ArtifactNet keys on this rather than on audible style cues.
Codec reconstruction error
Neural-codec decoders produce reconstruction residuals whose structure diverges from human recordings — amplified by a separation model trained only on real music.
Unseen-generator robustness
Because the signal is generation-mechanism physics, not generator-specific patterns, detection holds up on generators absent from the training set.
Use cases
Who uses it
Distributors
Catalog-scale artifact detection
Batch-process thousands of tracks via REST API. Flag AI-generated content with RVQ signatures before ingestion into your catalog. JSONL / CSV / XLSX results export.
Mastering
Pre-master cleanup
Insert de-artifact on the 2-bus before limiting. Seven factory presets tuned for different generator profiles — AI aggressive, lossy conservative, HPSS harmonic focus.
Platforms
Ingestion pipeline integration
REST API with batch create / append / commit flow. Async processing via worker queue. Results polled by job ID. Works within standard 10 MiB request limits via R2 presigned URLs.
Producers
Real-time DAW plug-in
VST3 / CLAP for macOS and Windows. 186 ms ECO latency for live monitoring, 500 ms Standard for final print. No cloud dependency — all inference on-device.
Benchmark
ArtifactBench results
ArtifactBench is a held-out evaluation set of 6,183 tracks spanning 22 AI music generators (Suno, Udio, Stable Audio, MusicGen, Riffusion, EnCodec, and 16 others) and 6 real music sources. Results are from the paper: Oh, H. (2026). ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics. arXiv:2604.16254.
0.9829
F1 score
vs. CLAM 0.7576 · SpecTTTra 0.7713
(identical eval conditions)
1.49%
False positive rate
Real music incorrectly flagged as AI
5–10 s
Inference per track
4-minute track on RTX 4090
OOD
Out-of-distribution robust
Generalizes to unseen generators not in training set
FAQ
Common questions
What AI music generators does ArtifactNet detect?
ArtifactNet detects AI-generated music from any generator that uses neural audio codecs with residual vector quantization (RVQ), including Suno, Udio, Stable Audio, MusicGen, Riffusion, and EnCodec-based systems. Because it targets the physics of RVQ residuals rather than generator-specific patterns, it also generalizes to unseen generators not present in training.
Is ArtifactNet a watermark detector?
No. ArtifactNet is not a watermark detector and does not require any cooperation from the generator. It analyzes forensic residual patterns — the physical artifacts that neural audio codecs inevitably leave behind — without any planted signal.
Is ArtifactNet open source?
Model weights and inference code are available on the Hugging Face repository under a CC BY-NC 4.0 license (non-commercial, personal, and research use). The paper is freely available at arXiv:2604.16254.
How accurate is ArtifactNet?
On ArtifactBench — a held-out set of 6,183 tracks across 22 AI generators and 6 real music sources — ArtifactNet achieves an F1 score of 0.9829 with a false positive rate of 1.49%, outperforming prior models including CLAM (F1 0.7576, FPR 69.26%) and SpecTTTra (F1 0.7713, FPR 19.43%) evaluated under identical conditions with published checkpoints.
Can I try ArtifactNet without signing up?
Yes. The free web demo at demo.intrect.io lets you upload any WAV, MP3, or FLAC file (or paste a YouTube URL) and get an AI-vs-human verdict instantly. No account required.
How do I use ArtifactNet at scale?
The ArtifactNet REST API (app.intrect.io) supports batch jobs — create a batch, append audio files, commit, and poll for results in JSONL, CSV, or XLSX format. The free tier includes 20 minutes of audio in total per account, with no credit card required.
API
Quick start
# 1. Upload tracks and start a batch job
curl -X POST https://api.intrect.io/v1/batch/create \
-H "Authorization: Bearer $TOKEN"
# 2. Append audio files
curl -X POST https://api.intrect.io/v1/batch/$BATCH_ID/files \
-H "Authorization: Bearer $TOKEN" \
-F "[email protected]"
# 3. Commit to queue
curl -X POST https://api.intrect.io/v1/batch/$BATCH_ID/commit \
-H "Authorization: Bearer $TOKEN"
# 4. Poll for results (JSONL / CSV / XLSX)
curl https://api.intrect.io/v1/batch/$BATCH_ID/results?format=jsonl \
-H "Authorization: Bearer $TOKEN"Pricing
Simple, usage-based plans
Free
$0
20 minutes total
- Batch API access
- JSONL / CSV export
- No credit card required
Start free →
Artist
$9.99/mo
100 minutes / month
- Everything in Free
- API key access
- Priority support
Upgrade →
Most popular
Creator
$49/mo
600 minutes / month
- Everything in Artist
- DDEX / Apple export
- Batch history 90 days
Get Creator →
Pro
$149/mo
2,000 minutes / month
- Everything in Creator
- Multiple API keys
- SLA support
Get Pro →
REST API
Batch processing
Upload hundreds of tracks, process asynchronously, download structured results. Free tier included — no credit card required to start.
Get API access →
DAW Plug-in
Real-time, on-device
VST3 / CLAP for macOS and Windows. 14-day full trial, no account required. Same ArtifactNet model, running entirely on your hardware.
Download trial →