Royalty Radar

Inspiration# Royalty Radar

Royalty Radar scans audio and video files for AI-generated music and cloned vocals, so artists, labels, streaming platforms, UGC creators and ad agencies can certify human work or flag AI. It never guesses.

Inspiration

AI can now copy an artist's voice and style in about a minute, and posting the result takes seconds. The original artist usually finds out late, if ever. Meanwhile, platforms, creators and ad agencies have no quick way to check what audio they are actually using. I'm a UGC creator myself and run an AI consulting brand on the side, so I sit on both sides of this problem at once — I post into the same ad and content pipelines where cloned vocals and AI tracks are already slipping through unchecked, and I've watched clients treat "is this AI?" as a shrug instead of a real answer. Royalty Radar is the tool I wished existed for both of those hats.

What it does

Royalty Radar takes audio from a file, a video, a link, or a live microphone recording, and it detects AI-generated vocals and music in both. For video files — ads, UGC clips, music videos — it automatically pulls the audio track out first, so an MP4 or MOV gets the same scrutiny as a standalone WAV or MP3, with nothing extra for the user to do. It sends that audio to HumanStandard's detector, and the verdict decides what happens next:

Verdict	What Royalty Radar does
human	Issues a certificate of human authorship, with the evidence the detector returned
ai at 80%+ confidence	Drafts a consent-and-revenue-share email to whoever posted it, written as an independent artist or as a label representative. It is a draft and is never sent automatically.
Anything else (uncertain, suspicious, or an ai verdict below 80% confidence)	Flags it for a person. It never guesses and never accuses on a shaky verdict.

Who it is for

The primary user is an independent artist, label or manager: someone who has no legal or trust-and-safety team and needs a quick, honest answer they can run themselves. The same tool helps:

• Streaming platforms, to screen uploads before they go live
• UGC creators, to show their audio is original
• Ad agencies, to check voiceovers and music — in the finished video file, not just an isolated audio clip — before an ad runs

How it runs on Quirq

Royalty Radar runs on Quirq's managed cloud at app.xo.builders. Claude Code runs inside an XO Space, follows written rules in CLAUDE.md, and scans a folder of files through scan_batch.py. For each file it reports the verdict, the confidence, what it produced, and why. The Space's Agents view shows the session, the project, the model, tokens and turns. The dashboard runs in the same Space, so the agent's scans and the dashboard's Results tab share one log.

How we built it

• HumanStandard integration: local files and links go to /api/analyze, the job is polled until complete, and the result fields (verdict, confidence, origin, evidence line) drive every output. Free mock scenarios let us test all the routing without spending credits.
• Video and ads: ffmpeg pulls the audio track out of videos automatically, so an MP4 or MOV can be scanned directly — the detector never has to know whether the source was audio-only or a video file.
• Dashboard (Streamlit): Check a track (upload or link), Listen (record from the microphone and scan), and Results (every scan and outcome).
• Listen tab: a small custom recorder captures audio in the browser and sends it over Streamlit's own connection. The built-in recorder was blocked by the workspace's login layer, so we built our own.
• Two email voices: a first-person independent artist, and a formal label representative. Both are drafts with placeholders for the sender to fill in.
• Stack: Python, requests, python-dotenv, Streamlit, ffmpeg, GitHub.

Real API results

These are real scans, run by the agent in the Space:

File	Verdict	Confidence	Output
Original song (human)	human	97.1%	Certificate. 25 of its 25 nearest reference recordings are verified human.
Real ad	human	99.7%	Certificate
AI ad with AI voiceover	ai	99.9%	Outreach draft
AI music video with AI song	ai	99.3%	Outreach draft
AI video with a human song	ai	7.2%	Manual review

Real job ID for the last one: 689959b3-ce96-4749-91e3-141b2a772736.

Safeguards and failure handling

• Scans are mock-first. The agent only runs real, credit-costing scans when asked.
• Outreach needs a confident ai verdict, and it is only ever drafted, never sent.
• Odd or missing data always routes to review instead of crashing.
• Scan polling retries brief network errors and fails fast on a bad key.
• The API key stays in the environment and is never printed or committed.
• 18 automated tests cover the routing, the emails, polling, and video handling.

Challenges

• Our own tool nearly accused a real musician. In a dry run, an AI video with a human-made song came back labeled ai, but the detector's own AI probability was 7.2%, and our first version turned that into an outreach email. We changed the routing so a low-confidence ai goes to a person with an explanation.
• Certificates can say more than the evidence supports. The detector returns human at about 97% for an instrumental track. It has no field that says whether vocals were present, so the certificate states what it rests on and does not claim more.
• The recorder was blocked. The hosted workspace's login layer intercepted the built-in recorder's upload, so we built a recorder that sends audio over the live connection.
• The Space's Sessions view trails the agent. It updates a little behind the run, so we show it after the run finishes.

Accomplishments

• One batch produces all three outcomes: two certificates, two outreach drafts, and one review flag.
• The tool refused to act on a contradictory verdict, and we could show why with the detector's own numbers.
• The whole pipeline, the agent, and the dashboard run from one Quirq Space.

What we learned

The verdict is not the product. What you do with it is. A detector that returns a label is only useful once the routing is honest about how sure it is, and once someone who would act on it can see the evidence.

What is next

• A vocal-presence check, so instrumentals are handled explicitly
• Watching feeds (an artist's tagged videos, ad libraries) instead of one file at a time
• A browser extension that captures a tab's audio for scanning
• Reporting the inconsistent verdict we found to HumanStandard
• A public landing page for the project

Tracks

humanstandard, the code registry, bring your own project, Quirq (Build It) Space ID: 826dcadd-8739-4cdc-a7c8-992dc7b6aa97, project: royalty-radar, Space name: royalty-radar

Built With

• humanstandard
• quirq