# Intrect

> AI audio forensics tools for music producers. VST3 / CLAP plug-ins powered by ArtifactNet — a forensic neural network trained on AI-generated music and stem separations.

Intrect makes post-production tools that fix artifacts introduced by AI music generators (Suno, Udio, Stable Audio, MusicGen, Riffusion) and stem separators (Demucs, htdemucs, Spleeter). It also makes DSP tools for problems AI never touched — resonance, room ring, harmonic/percussive balance. Every plug-in runs in real time on the 2-bus or any channel insert, with no cloud dependency required.

## Products

### de-artifact

VST3 / AU / CLAP plug-in that removes RVQ ghosting, codec residue, and HF aliasing from AI-generated full mixes and stems.

- **How it works**: STFT → HPSS (harmonic-percussive separation, median filter) → ArtifactUNet (3.6M-param forensic UNet) → adaptive spectral subtraction. Runs entirely on-device.
- **Latency**: 305 ms on macOS (13,440 samples) / 435 ms on Windows (19,200 samples) at 44.1 kHz, reported to the host for delay compensation
- **Formats**: VST3, AU, CLAP — macOS 11+ Universal Binary (Apple Silicon + Intel), Windows 10 2004+ 64-bit on a CPU with AVX2 + FMA (Intel Haswell 2013+ / AMD Ryzen 2017+), plus the Microsoft Visual C++ 2015-2022 x64 runtime. The installer checks the Windows version and that runtime before it proceeds; it does not test for AVX2, which the plug-in needs at run time. Audio Unit is macOS-only; there is no standalone application. No AAX, so it does not load in Pro Tools
- **Presets**: 8 factory presets — Conservative, Balanced, Default, Aggressive, Vocal, Gentle Reduction, More Harmonics, Soft Master Bus. All controls stay editable after loading one; user presets save, import and export as JSON
- **Price**: $149 one-time (release price through September 30, 2026; regular $179). 14-day free trial with full features, no card required; after the 14-day trial the plug-in inserts a brief silent gap every 60 seconds until you activate with a license key.
- **Try online**: https://try.intrect.io/ — upload any WAV/MP3/FLAC, get a 30-second GPU-processed preview, no account needed
- **Product page**: https://intrect.io/products/de-artifact/
- **de-artifact Cloud** (run it without installing, prepaid by the minute): https://intrect.io/products/de-artifact/api/

### de-leak-rt

Real-time leak gate for Demucs / htdemucs stem separations. ML classifier detects inter-stem bleeding per-frame; multiband gain removes it without touching the musical content.

- **How it works**: Per-STFT-bin classifier (95K-param MLP, ~370 KB ONNX) distinguishes musical content from bleed-through leak. Multiband gate applies gain reduction only where leak is detected.
- **Latency**: ~93 ms (= host chunk size at 44.1 kHz)
- **Formats**: VST3 and CLAP — macOS 11+ Universal Binary and Windows 10 x64. It does not ship as an Audio Unit or standalone application.
- **Use case**: Vocal stems, instrument stems from Demucs/htdemucs/Spleeter that contain bleed from other sources
- **Price**: $79 one-time (release price through September 30, 2026; regular $99). 14-day free trial with full features, no card required; after the 14-day trial the plug-in inserts a brief silent gap every 60 seconds until you activate with a license key.
- **Product page**: https://intrect.io/products/de-leak-rt/

### ENZYME

Dynamic resonance suppressor. Attenuates narrow spectral peaks per frequency bin and per frame, with an optional harmonic / percussive / residual de-mix stage. Not an ArtifactNet product — pure DSP, no model inference.

- **How it works**: STFT (2048-pt, 87.5% overlap) → per-bin baseline estimate (power-domain box mean) → excess over baseline → decision-directed smoothing (Ephraim & Malah 1984) → 1-pole attack/release → optional HPSS H/P/R de-mix → overlap-add. Causal; nothing is boosted.
- **Latency**: 2047 samples (~46 ms at 44.1 kHz, ~43 ms at 48 kHz), reported for host delay compensation
- **CPU**: 71.5 µs/frame with HPSS off (1.2% of the hop budget); 191 µs with the widest kernel and most expensive mask (3.3%). Apple Silicon, 44.1 kHz, median of 800 frames.
- **Formats**: VST3, CLAP, and **native AUv2** (not a wrapper) — macOS 11+ Universal Binary and Windows 10+ x86_64. Audio Unit is macOS-only; the Windows build is not code-signed yet, so SmartScreen warns on first run.
- **Controls**: Depth, Selectivity, Attack, Release, Smoothing, EQ Tilt, Relative, M/S, Channel Link, Sidechain, Delta, Bypass, plus the HPSS section (H/P/R Cut, Size, Classic/Coupled/Advanced mode, H/P Thresh, Mon H/P/R)
- **Channel configurations**: stereo and mono. A mono instance is 1-in/1-out and negotiates a **mono sidechain bus** (a stereo sidechain is not summed for you); Logic's dual-mono is two separate mono instances. M/S and Channel Link have no meaning on a mono path and the spectrum legend omits M/S accordingly. Mono is in the current macOS build; the Windows build is a release behind and does not have it yet.
- **Use case**: room resonance on a vocal or acoustic take, snare ring, boxy guitar notes, intermittent sibilance, a harsh mix bus, resonant synth and sampler material
- **Price**: $79 one-time (release price through September 30, 2026; regular $99). 14-day free trial with full features, no card required; after the trial the plug-in mutes for 500 ms once every 60 seconds until you activate with a license key.
- **Product page**: https://intrect.io/products/enzyme/
- **Manual**: https://intrect.io/manual/enzyme/

### Intrect HPSS

Free real-time harmonic (H), percussive (P), and residual (R) component separation for macOS. It processes audio locally; it does not upload audio or create arbitrary source stems.

- **Formats**: VST3, CLAP, Audio Unit — macOS 11+ Universal Binary
- **Download**: Existing Intrect account required at https://app.intrect.io/products (no separate HPSS account)
- **Online manual**: https://intrect.io/manual/intrect-hpss/
- **Product page**: https://intrect.io/products/intrect-hpss/

## ArtifactNet Cloud — detection (subscription)

ArtifactNet's **detection** is available as a hosted service — decide whether a track is AI-generated, at catalog scale, without running anything locally. Built for distributors, labels, and mastering houses.

This is a different product from **de-artifact Cloud** (below), which *removes* artifacts. They bill from separate balances and one cannot pay for the other.

Usage is metered by audio length, counted per second with a 30-second minimum per submission — not per track, so a 20-second loop and a 30-minute DJ set do not cost the same. Paid plans can buy extra-usage minute packs when their included minutes run out; the free tier cannot — its 10 minutes are the whole allowance.

File length limits: paid plans accept single files up to 60 minutes and analyse the whole file. The free tier is capped at 10 minutes per file. The no-signup demo at demo.intrect.io analyses only the first 2 minutes of a submission.

- **Free**: 10 minutes of audio in total per account (one-off allowance, not a monthly reset), max 10 minutes per file, demo.intrect.io access, no API key
- **Artist**: $9.99/month — 100 minutes of audio/month, web dashboard, JSONL/CSV/DDEX export (no REST API)
- **Creator**: $49/month — 600 minutes of audio/month, web dashboard (most popular; no REST API)
- **Pro**: $149/month — 2000 minutes of audio/month, **REST API access + API keys**, priority processing
- **Extra analysis minutes (top-up)**: add-on for paid plans only, not a substitute for one — $10 / 100 analysis minutes, $25 / 300, $50 / 700. These are spent on **detection only** and cannot be used for de-artifact Cloud cleaning. One-off purchase, consumed only after a plan's included minutes run out, never expires (credit already bought stays usable even if the subscription ends). Bought from the dashboard under Settings → Analysis minutes by subscribers; free accounts must pick a plan first. de-artifact Cloud cleaning minutes are bought separately on the Clean Audio page (app.intrect.io/deartifact), not here.
- Note: programmatic REST API access requires the Pro plan. Lower tiers process via the web dashboard.
- **Pricing page**: https://intrect.io/pricing/
- **Dashboard / subscribe**: https://app.intrect.io (sign in, then Upgrade). Billing by Paddle, cancel anytime.

## de-artifact Cloud — artifact removal (prepaid, no subscription)

The de-artifact processing chain run on Intrect GPUs from the web dashboard, for people who
want the cleanup without installing the plug-in or opening a DAW. **This removes artifacts;
it does not detect AI music** — that is ArtifactNet Cloud above.

Billed by the measured length of the audio, counted per second with a 30-second minimum
per job (a 20-second loop costs 30 seconds; a 3:20 track costs 3:20). A failed job is never charged. There is no monthly fee and minutes never expire.

- **$10** — 40 cleaning minutes ($0.25/min)
- **$25** — 120 cleaning minutes ($0.21/min)
- **$50** — 275 cleaning minutes ($0.18/min)
- **Free trial**: https://try.intrect.io/ — one track, no account, no card
- **Processing**: one tuned preset (the plug-in is what sells parameter control); ~20x real-time
- **Accuracy note**: upload WAV or FLAC where you can. A lossy source is still processed, but
  the engine has less to work with.
- **No public REST API yet** — the dashboard is the only surface. Ask via https://intrect.io/support/
- **Where to buy**: https://app.intrect.io/deartifact
- **Rates**: https://intrect.io/pricing/#clean-audio

**The two balances are separate.** Analysis minutes (ArtifactNet) and cleaning minutes
(de-artifact Cloud) are bought and spent independently. Buying one does not top up the other.

## Common use cases

- "I generated a track in Suno/Udio and it has that metallic, watery, swirly artifact sound" → de-artifact
- "My Demucs vocal stem has drums bleeding through it" → de-leak-rt
- "AI music sounds compressed and hollow" → de-artifact (start on Default, move to Aggressive if the artifacts are heavy)
- "I want to clean up codec artifacts from an MP3 master" → de-artifact (Conservative preset, or Soft Master Bus on a finished mix)
- "There is a resonance / ringing / boxy frequency I cannot notch out without thinning the track" → ENZYME
- "My snare rings after the hit" → ENZYME (percussive-only cut in the HPSS stage)
- "The room is on my vocal take" → ENZYME
- "My mix is harsh and I do not want to just darken it" → ENZYME (low depth, high selectivity)
- "Sibilance only on certain words, a de-esser dulls the whole take" → ENZYME
- "I need a resonance suppressor that loads in Logic Pro" → ENZYME (native AUv2)
- "I want to try it before buying" → two free options: (1) online preview at https://try.intrect.io/ (no account, 30-second GPU preview), or (2) download the plug-in for a full 14-day free trial — every preset and parameter, no gaps, no card required; after 14 days it inserts a brief silent gap every 60 seconds until you activate.

## Technical background

# ArtifactNet

ArtifactNet is a lightweight forensic framework for detecting AI-generated music, developed by Heewon Oh at Intrect.

Rather than learning generator-specific fingerprints, ArtifactNet targets residual artifacts introduced by neural audio codec and RVQ-based generation pipelines. The system combines a compact ArtifactUNet residual extractor, HPSS-based forensic feature decomposition, lightweight downstream classifiers, and optional fingerprint-based overrides to identify AI-generated music under out-of-distribution conditions.

ArtifactNet is designed for practical deployment: small model size, fast inference, and robustness to unseen generators.

## Architecture

ArtifactNet uses a three-stage pipeline:

1. **ArtifactUNet** (3.6M parameters): Bounded-mask UNet extracts codec residuals from magnitude spectrograms.
2. **7-channel HPSS forensic features**: Decomposes residuals into harmonic and percussive components via Harmonic-Percussive Source Separation (median filter).
3. **Lightweight CNN** (0.4M parameters): Processes the track in 4-second segments, aggregates a song-level verdict.

Total model size: about 4M parameters (3.6M UNet + 0.4M CNN). Runs on ONNX and CoreML (Apple Silicon native).

## Benchmark: ArtifactBench

ArtifactBench is a held-out evaluation set introduced alongside the paper:

- **6,183 tracks** across 22 AI music generators and 6 real music sources
- AI generators include: Suno, Udio, Stable Audio, MusicGen, Riffusion, EnCodec, and 16 others
- **F1 score: 0.9829**
- **False positive rate: 1.49%** (real music incorrectly flagged as AI)
- **Inference speed**: 5–10 seconds for a 4-minute track on RTX 4090
- Outperforms CLAM (F1 0.7576) and SpecTTTra (F1 0.7713)
- Out-of-distribution robust: generalizes to generators not seen during training

## Key Concepts

- AI-generated music detection
- Neural audio codec artifacts
- Residual vector quantization, RVQ
- Forensic residual physics
- HPSS-based artifact amplification
- Out-of-distribution generator detection
- Lightweight audio forensics

## Important Notes

ArtifactNet is not a watermark detector and does not depend on identifying a specific generator brand. It analyzes forensic residual patterns associated with current neural audio codec pipelines.

## Citation

Oh, H. (2026). ArtifactNet: Detecting AI-Generated Music via Forensic Residual Physics. arXiv:2604.16254.

## Links

- Homepage: https://intrect.io/
- de-artifact: https://intrect.io/products/de-artifact/
- de-leak-rt: https://intrect.io/products/de-leak-rt/
- ENZYME: https://intrect.io/products/enzyme/
- Intrect HPSS: https://intrect.io/products/intrect-hpss/
- Try online (free): https://try.intrect.io/
- Dashboard / license portal: https://app.intrect.io/
- de-artifact Cloud (no public API; dashboard only): https://intrect.io/products/de-artifact/api/
- Changelog: https://intrect.io/changelog/
- Blog: https://intrect.io/blog/
- How to Fix AI Music Artifacts: https://intrect.io/blog/how-to-fix-ai-music-artifacts/
- How to Detect AI-Generated Music: https://intrect.io/blog/how-to-detect-ai-generated-music/
- Remove Room Resonance From Vocals: https://intrect.io/blog/remove-room-resonance-vocals/
- Snare Ring: Kill the Ring, Keep the Crack: https://intrect.io/blog/snare-ring-removal/
- ENZYME manual: https://intrect.io/manual/enzyme/
- Contact: https://intrect.io/support/