Resemble AI

Tool · this deep dive is built from the official pages of the platform itself and of the tools that support it.

Overview

Resemble AI is a generative-AI security platform plus voice models. The product list is from the site's machine-readable index, llms.txt.

Detect — deepfake recognition in audio, video and images, including in calls and meetings. Verify — provenance watermarks: PerTh for audio, VideoSeal for images and video. Identity — registration and matching of a speaker from a sample of four seconds and up. The voice part sits in open models: Chatterbox (speech synthesis with zero-shot voice cloning), Chatterbox Turbo, Chatterbox Multilingual, DramaBox. The footer separately lists Resemble TTS, Resemble Voice Creation, Resemble Audio and Resemble STS.

Chatterbox Multilingual is declared as "20+ languages", with the list named in full on the model page; Russian, Arabic, Hindi and Chinese are in it. The model is under the MIT licence and does not depend on the plan. For deepfake recognition, validation in 51 languages is declared.

Which platforms it works with

The answer is known by half. Direct export of finished audio to publishing platforms is not in what was read: no section about export to social networks or streaming was found on the site.

Documented instead are workflow and security integrations: Zoom, Microsoft Teams, Google Meet and Webex for Resemble Meetings, a Chrome extension, SIP and SIPREC for telecom operators, SDKs for Python, Node.js and JavaScript, a page Integrations and environments. Deployment is declared in the cloud, on your own site and in an isolated perimeter.

Pricing

Prices from the pricing page, currency US dollars.

TEAM — 350 a month monthly, 280 with annual payment (yearly saving named as 840): 5 seats, reduced processing rates, Intelligence for meetings, large and batch uploads.

BUSINESS — 1000 monthly, 800 with annual payment (saving 2400 a year): 20 seats, company-wide calendar integration, security notifications for meetings, SSO.

ENTERPRISE — price by agreement: volume discounts, SLA, model fine-tuning, installation on your own site.

The minimum paid step is 280 a month, TEAM with annual payment. FLEX costs 0 and is charged per unit.

One reservation matters: the table covers Detect, Intelligence, Meetings, Identity and Watermarker. A separate price list for speech synthesis and voice cloning is not on the page.

What is free

FLEX — 0 dollars a month: no subscription fee, no card, payment by fact out of credits bought in advance. It includes deepfake recognition in audio, images and video, Intelligence analytics, identity registration, real-time recognition for meetings, SynthID checking and C2PA signing of recordings, platform and API access. Seats: one.

On the credits the page answers verbatim: "No, credits on the Flex Plan never expire". The zero is the absence of a subscription fee, not free work: processed seconds and calls are charged from the first one. The

Restrictions

Limits in minutes do not exist: charging is per second and per call, and the rate depends on the plan. Rates are from the "Processing rates" block of the pricing page:

Detect for audio and images — 0.03500 dollars per second on FLEX, 0.01500 on TEAM and BUSINESS. Detect for video — 0.07000 and 0.03000. Intelligence — 0.02500 and 0.01500. Identity search — 0.00050 per call. Watermarker: encode 0.00050, decode 0.00020 per call. ENTERPRISE has volume pricing.

Seats: 1 on FLEX, 5 on TEAM, 20 on BUSINESS. FLEX credits never expire, and there are no minimum commitments.

Rights to the output

The answer is known in part. The terms of service contain no transfer of rights to the generated audio: a wording about ownership of the Output is not in the text, and differences by plan are not described.

Resemble retains rights to Aggregated Data — anonymised derivatives and metadata — and bans using the output "to train, improve, or otherwise further develop your own or any third party's product".

For the open Chatterbox Multilingual it is declared "Full model weights on Hugging Face … No vendor lock-in, no commercial restrictions" under the MIT licence — which concerns the self-hosted model, not work through the platform.

What was lacking: who owns audio created through the platform does not follow from what was read.

Do they train on your data

The answer is known in part: the terms separate two notions. Content — what you uploaded — is used "for the purposes of having Resemble AI synthesizing and generating an AI voice model". Aggregated Data — anonymised derivatives and metadata — "may be used for … internal learning and training", and also for debugging, error correction and product development.

What was lacking: a direct statement that the original user recordings go into training of the product models is not in the text, and the extent of training on raw audio is not disclosed.

An opt-out is not described at all — not by a setting, not by a plan. But this is not a statement that opting out is possible.

How you earn with it

Commercial use is directly permitted in one branch — the open one: for Chatterbox Multilingual "no commercial restrictions" is declared under the MIT licence, independently of the plan.

For output created through the platform there is no answer: no transfer of rights, no differences by plan. The absence of an answer is neither permission nor prohibition. And one ban operates over everything: the output may not be used to develop your own product or a third party's.

What it will not do

The sharpest point here is whose voice may be uploaded.

The consent of the person whose voice is cloned is required. The terms of service say verbatim, with the spelling of the source preserved: "Resemble may require consent form the individual or third party whose voice is being cloned. Consent needs to be verbal, unless otherwise stated by Resemble". Consent by default is verbal, not written; another form only if Resemble states otherwise. The user also warrants that they own the content or hold "the necessary licenses, rights, consents and permissions". The ethics page puts it more strictly: "voice cloning only occurs with the explicit consent of the individual". Consent is checked: identity is matched by the Resemblyzer speaker-recognition technology.

A separate direct ban on cloning the voice of a public figure is not in what was read; the nearest is the general ban on "impersonation without authorization". This is not permission for public figures: the consent requirement has no exceptions in the source. Missing is only a separately declared norm about them.

Forbidden subjects are named as a list: hate speech and discrimination, terrorism and violence, exploitation of children, unattributed content and distortion of the source, unwanted mass mailings, misleading messages. The terms add false claims of affiliation and phishing for financial details.

Labelling is uneven. For the open models it is obligatory and automatic — "Every clip is watermarked with PerTh at generation, before it leaves the model" for Chatterbox Multilingual, and the same for every output of DramaBox. In the terms Resemble Watermarker is mentioned only as a product: obligatory labelling of all platform outputs, and differences by plan, are not described there.

Verified data

The checked data this deep dive rests on.

About the platform

A generative AI security platform plus voice models. Detect — recognising deepfakes in audio, video and images, including in calls and in Zoom, Teams, Meet and Webex. Verify — provenance watermarks: PerTh for audio, VideoSeal for images and video. Identity — enrolling and matching a speaker from a sample of four seconds or more. The voice side is carried by open models: Chatterbox (TTS with zero-shot voice cloning), Chatterbox Turbo, Chatterbox Multilingual, DramaBox; the site footer separately lists Resemble TTS, Resemble Voice Creation, Resemble Audio and Resemble STS

llms.txt is the official machine-readable index of the site, declared in Resemble's own robots.txt; the product set was checked against the footer of https://www.resemble.ai/pricing

source, checked 2026-08-07

Chatterbox Multilingual: «20+ languages» (verbatim) — Arabic, Chinese, Czech, Dutch, English, Finnish, French, German, Hebrew, Hindi, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Turkish, Vietnamese. Russian, Arabic, Hindi and Chinese are in the list. The model is under the MIT licence and does not depend on the plan. For deepfake detection, validation in 51 languages is claimed

The figure of 51 languages for Detect is taken from https://www.resemble.ai/llms.txt

source, checked 2026-08-07

Platforms

There is no direct export of finished audio to publishing platforms in what was read. Workflow and security integrations are documented: Zoom, Microsoft Teams, Google Meet and Webex for Resemble Meetings, a Chrome extension, SIP and SIPREC for telecoms operators, SDKs for Python, Node.js and JavaScript, and a separate Integrations and environments page. Deployment is claimed in the cloud, on premises and in an air-gapped environment

There is no section on exporting the result to social networks or streaming on the site; only workflow integrations are listed

source, checked 2026-08-07

Pricing

FLEX — $0 a month, no subscription fee and no card, pay-as-you-go from credits bought in advance. Included: deepfake detection in audio, images and video, Intelligence analytics, identity enrolment, real-time detection for meetings, SynthID verification and C2PA signing of recordings, access to the platform and API; 1 team seat. «No, credits on the Flex Plan never expire» (verbatim)

The composition of FLEX and the answer about credits are taken from the pricing table and the FAQ block of the same page

source, checked 2026-08-07

TEAM — $350 a month paid monthly and $280 a month paid yearly (a saving of $840 a year), 5 seats, reduced processing rates, Intelligence for meetings, large and batch uploads. BUSINESS — $1,000 a month monthly and $800 a month yearly (a saving of $2,400 a year), 20 seats, company-wide calendar integration, security alerts for meetings, SSO. ENTERPRISE — price by agreement (volume discounts, SLA, model fine-tuning, on-premises installation). IMPORTANT: the price table covers Detect, Intelligence, Meetings, Identity and Watermarker; there is no separate price list for speech synthesis and voice cloning on the pricing page

The page title is «Deepfake Detection & AI Security Pricing | Resemble AI», which confirms the current positioning of the price list

source, checked 2026-08-07

280 USD a month — the TEAM plan paid yearly; paid monthly 350 USD a month. The free FLEX tier costs 0 USD and is billed per unit

The amounts are taken from the pricing table and from the «Transparent flexible pricing» summary table on the same page

source, checked 2026-08-07

Limits and restrictions

There are no limits in minutes; billing is per second and per call. Detect for audio and images: $0.03500 a second on FLEX and $0.01500 on TEAM and BUSINESS. Detect for video: $0.07000 and $0.03000 a second. Intelligence: $0.02500 and $0.01500 a second. Identity search: $0.00050 a call. Watermarker encode — $0.00050, decode — $0.00020 a call. For ENTERPRISE volume pricing is claimed. Seats: 1 on FLEX, 5 on TEAM, 20 on BUSINESS. FLEX credits never expire and there is no minimum commitment

The rates are taken from the «Processing rates» block of the summary table

source, checked 2026-08-07

Restrictions

WHOSE VOICE: consent is required from the person whose voice is cloned. Terms of Service: «Resemble may require consent form the individual or third party whose voice is being cloned. Consent needs to be verbal, unless otherwise stated by Resemble» (verbatim, the source's spelling preserved) — that is, consent is VERBAL by default, not written. The user also warrants that they own the content or hold «the necessary licenses, rights, consents and permissions». The ethics page: «voice cloning only occurs with the explicit consent of the individual» (verbatim). THE CHECK: identity when consent is obtained is verified by Resemble's Resemblyzer speaker recognition technology. There is NO separate outright ban on cloning the voice of a public figure — no such provision occurs in what was read; the nearest is the general ban on «impersonation without authorization». FORBIDDEN TOPICS: hate speech and discrimination, terrorism and violence, exploitation of children, unattributed content and misrepresentation of the source, unsolicited bulk messaging, misleading communications; the terms additionally forbid false claims of affiliation and phishing for financial information

The wording on explicit consent and verification through Resemblyzer and the list of forbidden uses come from https://www.resemble.ai/company/ethics

source, checked 2026-08-07

Legal

The terms do NOT contain a direct transfer of rights in the generated audio to the user: there is no wording about ownership of Output in the text. Resemble reserves rights in Aggregated Data — anonymised derivatives and metadata. The user is forbidden to use the output «to train, improve, or otherwise further develop your own or any third party's product» (verbatim). No difference by plan is described in the terms. A separate branch is the open models: for Chatterbox Multilingual it is claimed that «Full model weights on Hugging Face ... No vendor lock-in, no commercial restrictions» under the MIT licence

The caveat about MIT and the absence of commercial restrictions applies only to the self-hosted open model; the source is https://www.resemble.ai/learn/models/chatterbox-multilingual

source, checked 2026-08-07

The terms permit internal training on derived data: Content is used «for the purposes of having Resemble AI synthesizing and generating an AI voice model» (verbatim), and Aggregated Data «may be used for ... internal learning and training» (verbatim), as well as for debugging, fixing errors and product development. There is no direct statement in the text that the original user recordings go into training the product models

The terms do distinguish between Content and Aggregated Data, but the extent of training on raw audio is not disclosed

source, checked 2026-08-07

The PerTh watermark is Resemble's own technology; for the open models labelling is claimed to be mandatory and automatic: «Every clip is watermarked with PerTh at generation, before it leaves the model» (verbatim, Chatterbox Multilingual); for DramaBox — «Every output watermarked with PerTh at generation». The mark is claimed to be inaudible and resistant to re-encoding. In the Terms of Service the Resemble Watermarker is mentioned only as a product: mandatory labelling of all platform outputs and any difference by plan are not described there. Watermarker encode and decode calls are billed

The statement about mandatory labelling concerns specific models; no platform-wide rule was found in the terms

source, checked 2026-08-07

Catalogue section: all similar See also: catalogue index · scheduler comparison · find by situation · platform restrictions