Media, Gaming & Sports
Copyright Similarity Detection API
Finding a likely copy is the easy half. Deciding whether it is an infringement is a judgement about licences, context and law — and that part should never be made by a score.
- User-generated content platforms
- Music catalogues
- Publishers
- Studios and rights administrators
- Education and assessment providers
What your customers would see
example interfaceMockupIllustrative mockup of the interface we would build. Figures shown are examples, not results from a live deployment.
The problem this solves
Any platform that accepts uploads eventually needs to answer the same question at scale: has this file appeared before, and does it match something somebody owns. Manual review does not survive volume, so the matching has to be automated. The temptation is then to automate the consequence as well, and that is where platforms get into trouble.
Similarity is a measurement. Infringement is a legal conclusion that depends on ownership, licence terms, territory, duration, transformation, quotation and fair-dealing exceptions that differ by jurisdiction. A fifteen-second musical match might be a licensed sync, a permitted quotation in criticism, a coincidental chord progression, or a straight rip. The audio fingerprint cannot distinguish between those, and a system that acts as if it can will take down lawful work.
So the engineering goal is different from what it first appears. Not a verdict, but good evidence: where the match occurs, how long it runs, how strong it is, what kind of transformation sits between the two, and what a reviewer needs in order to decide quickly. Precision matters more than recall here, because every false positive lands on a real person.
Who buys this
User-generated content platforms, Music catalogues, Publishers, Studios and rights administrators, Education and assessment providers
What the API returns
What comes back
- Match candidates with similarity scores and the segments that matched
- Time-aligned match regions for audio and video, with start and end offsets
- Transformation indicators: pitch shift, time stretch, crop, re-encode, overlay, translation
- Text similarity with passage-level alignment and quotation-length measurement
- Perceptual image matching resilient to resize, crop and recompression
- A reviewer evidence bundle: side-by-side segments and the reasoning behind the score
- Confidence bands and explicit no-decision outcomes where evidence is weak
- Full audit record of every comparison, reproducible later
Endpoints
| Method | Path | Returns |
|---|---|---|
| POST | /v1/match/audio | Match an audio file against a reference catalogue |
| POST | /v1/match/video | Match video, with audio and visual tracks scored separately |
| POST | /v1/match/text | Passage-level text similarity against a corpus |
| POST | /v1/match/image | Perceptual image matching |
| POST | /v1/references | Register or update reference works in the catalogue |
| GET | /v1/matches/{id}/evidence | Reviewer evidence bundle for one match |
Request and response
illustrativeRequest
{
"asset": { "id": "upl_55210", "type": "audio", "duration_s": 212 },
"reference_catalogue": "label_core",
"min_segment_s": 7,
"return_evidence": true
}Response
{
"asset_id": "upl_55210",
"compared_at": "2026-10-22T10:02:41Z",
"engine_version": "sim-audio-2026.08",
"outcome": "review_required",
"matches": [
{
"reference_id": "ref_ttl_3317",
"reference_title": "Harbour Lights (studio master)",
"similarity": 0.91,
"matched_duration_s": 24.8,
"asset_region": { "from_s": 61.2, "to_s": 86.0 },
"reference_region": { "from_s": 12.4, "to_s": 37.2 },
"transformations": ["pitch_shift_+1.5st", "time_stretch_1.02x"],
"confidence": "high"
},
{
"reference_id": "ref_ttl_8802",
"similarity": 0.44,
"matched_duration_s": 8.1,
"transformations": [],
"confidence": "low",
"note": "Short match on a common four-chord progression. Not evidence of copying."
}
],
"decision": null,
"decision_note": "This API returns similarity evidence only. Infringement determination requires human review of ownership, licence and context.",
"audit_reference": "cmp_2026102210024100912"
}Where the data comes from
| Your reference catalogue | The works you own or administer, fingerprinted on ingestion — the quality of this set bounds everything else |
| Licensing and ownership records | Who owns what, in which territory, for how long; without this a match cannot be assessed at all |
| Existing licence and clearance data | Sync licences, blanket agreements and permissions already granted, so legitimate uses are not flagged repeatedly |
| Public-domain and open-licence registers | So works that are free to use are not pushed into a review queue |
| Reviewer decision history | Confirmed outcomes from your own moderators, which is the only honest way to calibrate thresholds |
| Third-party fingerprint services | Where you already licence one, matched alongside rather than replaced — their terms generally forbid redistributing reference fingerprints |
The intelligence layer
The difference between a demo and a service is mostly in this section. These are the parts that keep the answer trustworthy once real traffic is hitting it.
- Audio fingerprinting on spectral landmarks, which survives re-encoding, pitch shift and moderate time stretch
- Visual matching on frame-level perceptual hashes plus temporal alignment, so a cropped or re-timed clip still aligns
- Text matching combining shingled exact matching with embedding similarity, because paraphrase and verbatim copying need different detectors
- Segment-level alignment rather than whole-file scoring, since partial use is the normal case and a file-level score hides it
- Explicit minimum-duration and commonality filters so short, generic or widely shared fragments do not generate matches
- Threshold calibration against your reviewers' confirmed decisions, reporting precision and recall separately at each operating point
- Deliberate precision bias: the cost of a false positive here is a wrongly removed lawful work, which is not symmetric with a missed match
- Versioned engines with stored inputs, so a comparison made months ago can be reproduced exactly as it was made
What it changes for the business
Review capacity
Moderators spend their time on the ambiguous cases instead of re-finding obvious duplicates.
Fewer wrongful removals
Evidence bundles and precision-biased thresholds reduce the takedowns that generate appeals and press.
Defensible process
A reproducible audit record per comparison, which is what a dispute or a regulator asks to see.
Catalogue control
Rights holders get a clear picture of where their works appear without the platform having to guess at consequences.
How it fits into your product
How it embeds
- Asynchronous matching on upload with webhook callbacks, since fingerprinting a long video is not a request-time operation
- Match results routed into your existing moderation queue with the evidence bundle attached
- Reference catalogue managed through the API so new works are protected from registration onwards
- Deployment inside your own infrastructure, which is usual when the reference catalogue is commercially sensitive
- Configurable thresholds and minimum match durations set by your rights and legal teams rather than by us
- Appeal support: every match retrievable by reference for as long as your retention policy requires
Included in every build
- Authentication, scoped API keys and per-customer rate limiting
- Input validation with predictable, documented error responses
- Versioned endpoints so your integration never breaks without notice
- Request and response logging for audit and dispute resolution
- Latency, error-rate and usage monitoring with alerting
- OpenAPI specification, sandbox environment and client examples
- Backups, failover design and a documented recovery procedure
- An agreed support arrangement once the service is live
How we would build it
about 20 weeksDiscovery and feasibility
Weeks 1–3
This one depends heavily on data access, so we start by proving the sources exist, can be licensed for your use, and actually carry the signal the model needs. If they do not, you find out here.
Architecture and baseline model
Weeks 4–6
A written architecture and a baseline model measured against a holdout period, so you know what accuracy you are buying before the full build starts.
Production build
Weeks 7–16
Data pipelines, model training and serving, historical archive, validation harness, API, security, monitoring and documentation.
Validation and integration
Weeks 17–20
Backtesting against held-out history, accuracy reporting your team can audit, integration support and an agreed retraining cadence.
Commercial models
There is no fixed price list, because the cost is driven by data complexity, integration count, model requirements, request volume and the support level you need. Scope is agreed in writing before development starts.
Fixed-scope prototype
A defined slice of the service, built and priced as one piece of work. The usual first step when the business case needs proving internally.
Fixed-price build
An agreed scope against a written architecture, delivered for an agreed price. Best when the requirements are already clear.
Build then retainer
We build the service, then operate and improve it under a monthly agreement covering monitoring, updates and model maintenance.
White-label product
The service runs under your brand, your domain and your pricing, with MLPivot invisible to your customers.
Questions buyers ask
Will it tell us whether something infringes?
No. It returns similarity evidence — what matched, where, how strongly, and through what transformation. Infringement depends on ownership, licence, territory, duration, transformation and jurisdiction-specific exceptions such as fair dealing or quotation. Those are legal judgements. The response carries an explicit null decision for that reason, and we would resist any design that fills it in automatically.
Can we wire it straight to automated takedown?
We would advise against it and would want that written down before the build. Acting automatically on a raw similarity score causes real harm: licensed uses, criticism, parody, education and coincidental similarity all score above threshold. A sensible design automates the queue, not the consequence — and keeps a human decision and an appeal route in front of any removal.
How well does it handle deliberate evasion?
Pitch shift, time stretch, re-encoding, crop and overlay are handled routinely. Heavier transformation — live re-performance, substantial re-editing, or a translated text rewritten in different words — degrades the score, and sometimes below any usable threshold. The API reports low confidence rather than guessing, and we would rather miss a match than manufacture one.
What about short or generic fragments?
They are filtered out on purpose. A common drum pattern, a stock chord progression or a three-word phrase will match many works and mean nothing. Minimum duration and commonality filters are configurable, and the sample response deliberately shows such a candidate returned with low confidence and a note rather than hidden.
Can you match against a catalogue we do not own?
Only with the rights holder's agreement. Fingerprinting a third-party catalogue to match against requires a commercial arrangement with whoever administers it, and existing fingerprint vendors generally forbid redistributing their reference data. We identify that during discovery rather than after you have built a workflow that assumes it.
Related blueprints
Video Metadata Intelligence API
Scene detection, subtitles, topics, chapters and summaries generated from video files, returned as structured metadata your catalogue can store.
Content Recommendation API
Personalised rankings, related-item lists and discovery rows for streaming services and publishers, delivered as an API your product calls.
Audience Intelligence API
Interest profiles, consumption trends and segment analysis computed from your own audience data and returned as an API your product teams can query.
Want this built for your platform?
Tell us what your product does today and what you want another system to be able to ask it. We will come back with an approach, a scope and a price.