Resemble AI combines deepfake detection with media provenance

Resemble AI analyzes media for signs of AI generation or manipulation and provides tools for watermarking and identity verification. Its current commercial focus is security and provenance. The company has stopped selling voice generation, so a buyer arriving from an older voice-cloning comparison should begin with the current detection offer rather than an archived synthesis tutorial.

01 / ProductA detection platform with a clear product transition

The product notice explicitly ends commercial voice generation while retaining voice models as open research resources. This is a material product boundary, not simply a change in homepage language. Resemble remains active, but a production plan for paid text-to-speech should be evaluated elsewhere. Its history of generating speech helps explain its current direction without establishing that every synthetic voice can be detected.

Resemble Detect covers audio, images and video, while Intelligence supplies explanations and contextual analysis. The company promotes cloud API and on-premises deployment. A detection score, an explanation and an operational decision are different objects: the first estimates a property of the submitted material, the second describes supporting signals, and the third determines what a reviewer should do.

Watermarker addresses provenance from the other direction. A publisher marks an asset and later checks for that mark. This can complement detection on media of unknown origin, but an absent watermark does not establish that a recording is authentic. The product page distinguishes open-source audio PerTh from the enterprise multimodal offering; licensing and deployment should follow that distinction.

02 / AudienceTeams that need evidence about media

A trust-and-safety team assessing uploaded videos has a different job from a contact center screening a live caller. The first can wait for a completed analysis and preserve a case record. The second must decide what an agent does while the conversation continues. Resemble documents both persisted file analysis and streaming audio, so the integration should follow the decision's timing rather than treating every mode as interchangeable.

A useful initial audience is a team already able to review suspicious material. Detection can help prioritize that queue, locate segments and add structured evidence. It is a weaker fit for a workflow that expects a single automated label to establish a person's identity, honesty or entitlement. Authentic recordings can carry false claims, and legitimate synthetic media can be clearly disclosed.

If the job is extracting spoken words, Deepgram's transcription workflow is an adjacent capability rather than the same analysis. If the job is producing narration, ElevenLabs addresses generation. Defining which question needs answering prevents a broad phrase such as voice AI from hiding fundamentally different products.

03 / WorkflowA proposed review queue for submitted media

Consider a proposed pilot for a publisher receiving short audio and video submissions. Start with material the team is authorized to analyze and independently knows how to categorize. Include ordinary recordings, disclosed generated clips and deliberately edited samples. Preserve the original files and record what happened to each copy before submission; otherwise compression or a shortened excerpt can become an unexplained difference between results.

The quick start submits media, saves a job identifier and retrieves the result after processing. The application must distinguish completed analysis from failed processing. A failed job should enter an operational retry or manual-review state, never silently become a clean bill of authenticity. Associate the identifier with the original submission so a reviewer can trace which asset was inspected.

The detection guide also describes secure upload tokens for media that should not be hosted publicly. It separates watermark findings from the deepfake verdict, and notes that streaming results are not persisted as Detect jobs. That changes the case-management design: a live integration must deliberately preserve the evidence it needs within its approved retention rules.

Present a short review card containing the media, completed findings, relevant timestamps and the human disposition. Ask reviewers to explain disagreement with a model result. Measure missed suspicious content and unnecessary escalations separately, then assess whether the queue is more useful. This is an evaluation design, not a claim that Sequenced has tested detection accuracy.

04 / PricingSubscription access and processing are separate costs

The pricing page offers Flex with no subscription fee and paid processing, alongside Team, Business and custom Enterprise arrangements. The table below uses displayed dollar amounts for the published monthly billing option, rather than the lower annual-equivalent figures. Confirm invoicing currency, applicable tax and contracted usage terms before purchase.

A subscription should not be interpreted as unlimited analysis. Audio and video use time-based processing units, images use an image unit, and additional analysis operations have their own rates. For a cost estimate, count the material actually submitted, including retries and separate analysis stages. A one-minute final clip may incur more than one operation if the workflow also asks for contextual analysis or watermark decoding.

The practical purchasing question is whether volume and access requirements justify a paid plan. Compare seats, batch uploads, identity integration and deployment requirements with the cost of processing the expected workload. On-premises deployment is an enterprise discussion; it should not be assumed available merely because the public API can perform the same analytical task.

05 / DistinctionsDetection and provenance answer different questions

Resemble's useful distinction is the ability to combine analysis of unknown content with marks applied to content the organization controls. A newsroom can inspect an incoming clip while a production team marks its own published media. Those workflows meet at the review record, but they do not provide the same type of evidence.

The detection documentation explicitly keeps watermark evidence separate from the deepfake verdict. That is a helpful interface principle: the presence of a known mark describes origin information, while detection estimates manipulation or generation. Preserve both instead of collapsing them into an unexplained traffic-light badge.

The company also advertises explanations, heatmaps and model benchmarks. Treat these as aids to examination and vendor-reported performance, not a guarantee about a new recording environment. An explanation may make a result easier to investigate without making its conclusion correct. The most useful evaluation therefore retains difficult examples, legitimate edits and cases where the available media is too limited for a confident decision.

06 / QuestionsResolve the response semantics before deployment

Which result is actionable? Agree how the application handles uncertain findings, partial visual or audio analysis, and unavailable secondary analysis. The docs describe independently asynchronous components, so a completed core result is not a reason to assume every explanatory field is final. The reviewer should be able to tell which operations actually ran.

What will be retained? Cloud, zero-retention and on-premises choices imply different operational arrangements. Ask for the exact scope that applies to the chosen service and contract. Retaining a source file in the organization's own case system is a separate decision from retention by Resemble; one setting does not resolve both responsibilities.

How does the model behave on the actual distribution channel? Test the codecs, recording devices, language mix and legitimate postproduction operations the team encounters. Use a held-out set when adjusting escalation rules, so the final check is not merely a replay of examples that shaped the policy. Revisit the acceptance decision after a meaningful model or pipeline change rather than presuming an earlier result transfers unchanged.

07 / DecisionChoose the question before choosing the interface

Resemble is worth examining when synthetic-media risk is already attached to an identifiable review process. Begin with one media type, one responsible reviewer and a defined action following a finding. Expand only once the team can explain both useful detections and false alarms. The purchase is more defensible when evidence improves a concrete workflow than when it merely adds an impressive score to every uploaded file.