AI Transparency Statement

Audexum makes synthetic speech. Everything the Service produces as audio is generated by a machine, not recorded by a person. We mark that audio so a machine can detect it: generated files carry an AudioSeal watermark in the audio signal itself — with one documented exception in section 4 — and our production service is configured to refuse to deliver audio it could not watermark.

Our watermark is our obligation, not yours. If you publish synthetic audio — above all a cloned voice of a real person — you have your own duty to tell your audience it is artificial. Our mark does not do that for you.

We do not train models on your inputs or outputs. We do not identify or match speakers, categorise people by biometric data, or infer emotions. The numbered sections below are the binding statement.

This is the transparency information that Audexum publishes under Article 50 of Regulation (EU) 2024/1689 (the EU AI Act), whose transparency obligations have applied since 2 August 2026. It covers audexum.com, the Studio, the Android and Windows apps, our public API, and the ElevenLabs-compatible API. It is part of our Terms; how we handle personal data is in the Privacy Policy, and what you may and may not generate is in the Acceptable Use Policy.

It describes what the Service does today. Where something is not yet done, this statement says so rather than rounding up.

All of the following output is produced by AI models. None of it is a human recording, a human transcriptionist, or a human support agent.

Audio the Service delivers to you is marked in two independent ways:

• Signal watermark. We apply AudioSeal (Meta, MIT-licensed), a watermark carried in the audio signal itself. It survives the MP3 file we deliver and conversion to other formats. Every further lossy re-encode (for example a re-upload to a platform that compresses audio again) weakens it, and after a second MP3 encode it can fall below the detection threshold. This is the machine-readable marking Article 50(2) asks for.
• What the watermark carries. The AudioSeal signal holds a 16-bit payload. We set the first 8 bits to a fixed value identifying Audexum as the origin, and the second 8 bits to one of 255 buckets derived from the account that requested the generation. Audio generated without an account carries bucket zero. Lossy compression can flip some of these bits even when the watermark itself is still detected, so we do not rely on the signal payload alone: the same bucket is also in the signed label below.
  The bucket is not an account identifier and cannot be turned into one. Roughly one in every 255 accounts shares any given bucket, so on its own it narrows a file to a large group and no further. It is computed with a secret key we hold, so a copy of the audio does not let anyone else work out — or even guess at — which account a file came from. We use it only alongside our own generation records, and only to investigate a specific report of abuse or of a file being passed off as a real recording. We do not use it for advertising, profiling or analytics, and we do not disclose it to third parties except where law compels us.
• Signed label. Every file carries a metadata comment (RIFF INFO in WAV, ID3 in MP3, Vorbis comment in Ogg) reading AI-generated by Audexum (audexum.com). Synthetic speech, not a human recording. followed by an Ed25519 signature over the generation date and the watermark payload. Anyone can verify it with the public key at /.well-known/audexum-provenance.json. It proves a file came from us while the metadata is intact. It is easy to strip, so a missing label proves nothing.

The production service is configured fail-closed. If the watermarking service is unavailable, we do not fall back to delivering unmarked audio: the request returns an error and the background job is marked failed. You may lose a generation when watermarking is down; you will not receive un-watermarked audio from the paths listed in section 4. If a failure of this kind cost you credits, tell us and we will put them back — see /refunds.

Watermarking is a technical measure, not a guarantee. A determined adversary with enough re-processing can degrade any audio watermark. We do not claim ours is unremovable.

Watermarking is applied on these delivery paths:

• Studio synthesis and the public synthesis API;
• the ElevenLabs-compatible API, which routes through the same handler;
• batch synthesis;
• synthesis with a custom cloned voice;
• audiobook chapters;
• dubbing output;
• the free public tools on the site.

The Kokoro voice preview at GET /api/voices/kokoro/{voice_id}/preview was the one exception; since 5 October 2026 it goes through the same watermarking step as everything else.

Detecting it yourself. Our free checker at audexum.com/detect reads both marks from any file, with no account. AudioSeal is also open source, so you can run its detector yourself. A watermark on its own means this audio carries an AudioSeal watermark; it does not by itself prove the audio came from Audexum rather than from another service using AudioSeal. A valid signed label does.

Asking us. Because of that limit, we offer a verification route. Send a clip to [email protected] with the subject line watermark verification and we will run our detector over it and tell you whether it is watermarked. Where you are the person whose voice is in the clip, or you are acting under a legal process, we will also check our records for a matching generation and tell you what we find. We aim to reply within 30 days. We are a small operation and this is checked by hand, so we give a period we can actually keep rather than a faster one we cannot.

What we can and cannot say: we can tell you whether a watermark is present, and whether our records show a matching generation. We cannot certify that an unmarked clip was not made by an AI, and a negative result is not evidence that a recording is genuine.

The assistant on our website, Aria, is an AI system. It is not a person, and no human is reading or answering your chat messages as you type them. Its replies are generated by a language model running on our own servers in Bulgaria; they can be wrong, and nothing it says varies our Terms, /refunds or prices. Do not send it confidential information or personal data about other people. What we keep from a chat, and for how long, is in the Privacy Policy.

Audexum offers custom voice cloning. To enrol a voice we require the speaker to read a consent sentence aloud, and we check by speech recognition that the sentence was spoken. We keep the resulting record together with the voice.

Be clear about what that check is worth. It shows that someone spoke the sentence into the recording. It does not prove identity, and it is not proof that the speaker in the enrolment audio is the person who consented. We do not treat it as proof. The responsibility for holding the rights and permissions to a voice you enrol stays with you, under the Terms and the Acceptable Use Policy.

Article 50(4) puts a duty on the person who deploys the output, not only on us. If you publish audio that imitates a real person — a cloned voice, a dubbed clip in a recognisable person’s voice, an edit that makes a real person appear to say something — that is a deep fake, and you must disclose to your audience that it is artificially generated or manipulated. Our watermark does not discharge that duty: it is machine readable, and the duty is to inform people.

Where you use Audexum inside your own product, these deployer duties are yours towards your own users. We cannot see how you present the audio and cannot disclose on your behalf.

The Service can host a generated clip on a public share page. Those pages are our own publishing, so the disclosure on them is ours to make: share pages carry a visible AI-generated label next to the player. If you have shared a clip that could mislead someone, say so where you post it and do not wait for us.

Share links are public to anyone who has the URL, and we have no expiry for them at present — see the retention table in the Privacy Policy. Delete a share you no longer want public.

If you believe audio generated or hosted by Audexum impersonates you or someone else, or clones a voice without permission, report it at /report. You do not need an account to report, and you do not need to identify yourself as the rights-holder to get the report looked at.

You can also email [email protected]. Give us the share link or the audio, and what is wrong with it. We will tell you what we decided and why. A voice cloned without the speaker’s permission is a breach of the Acceptable Use Policy and we remove it.

• No training on your content. We do not use your inputs, your uploads or the audio we generate for you to train, fine-tune or evaluate any model.
• No speaker identification or matching. We use a voice embedding to synthesise speech. We do not compare it against other speakers, search for a speaker across accounts, or try to work out who someone is from their voice.
• No biometric categorisation. We do not infer age, sex, ethnicity, health, political or religious characteristics from a voice.
• No emotion recognition. We do not analyse a recording to infer the speaker’s mood or emotional state.
• No real-time remote biometric identification, and no facial or video recognition of any kind. The media safety checks in section 12A classify content; they do not recognise anyone.

These are commitments about what the Service does, and we will update this statement before any of them changes. The practices prohibited by Article 5 of the AI Act are not part of what we build or sell.

Image and video generation runs on hosted third-party models; section 12A lists them.

Audexum runs third-party open models on our own hardware in Bulgaria, and hosted models through OpenRouter for the V4 voice tier and when our own GPUs are unavailable (see /subprocessors). Required third-party software notices are at /attributions. Which providers are involved in running the Service is at /subprocessors.

Our image and video tools are sold under our own product names, but the models behind them were developed by other companies and are run for us by the providers below. We did not build or train them. The provider of each engine is also shown on the activation screen before any of your data reaches it, and is listed at /subprocessors. We may change the engine behind a product name; when we do, this table changes first.

How images and videos are marked. Every result is marked in three ways, so that both people and machines can tell it is AI-generated:

• The provider's own marks are kept. Some models embed their own provenance signals, such as Content Credentials (C2PA), an invisible watermark or AI-generated metadata. We never re-encode the files we deliver, so whatever marks the model embedded reach you intact, and when a file carries Content Credentials we do not add our own metadata to it, because any change to the file would make its credential read as tampered. One exception: when you add narration to a video, the picture is copied unchanged but the sound is re-encoded into a new file, so container-level marks from the provider (such as a C2PA record or AI-generated metadata) may not survive in the narrated copy. Pixel watermarks and our own metadata remain, and your original clip is untouched.
• Our metadata. We add XMP metadata to every file that does not already carry Content Credentials: the IPTC digital source type trainedAlgorithmicMedia, which is the standard code for media created by an AI model, and the creator Audexum (AI-generated). Like any metadata it can be stripped by other tools, and removing it breaks our Acceptable Use Policy.
• A visible label. Every image and video carries an "AI-generated" label wherever we display it: in the app, in your library and on share pages.

Our metadata is a label, not a cryptographically signed manifest. We will move to signed Content Credentials if the AI Act code of practice requires it. When you publish a realistic result, the disclosure duty in section 8 applies to you, exactly as for audio.

Safety checks. We check every prompt and every result (each image, and three frames of each video) with an automated classifier, a Google model reached through OpenRouter. Uploaded images are checked by the provider's own filters, not by ours. The providers also run their own checks. All of this is to block content our Acceptable Use Policy forbids. Those checks classify content, for example whether an image is sexual; they never identify who a person in it is. Blocked generations cost you nothing.

Providers' usage policies. Section 13B of the Acceptable Use Policy binds you to the usage policies of the providers and developers above, as they apply to the engine you use:

• fal: Acceptable Use Policy, fal.ai/legal/acceptable-use-policy (all fal engines);
• Black Forest Labs: Usage Policy, bfl.ai/legal/usage-policy (Image);
• Google: Generative AI Prohibited Use Policy, policies.google.com/terms/generative-ai/use-policy (Image Pro);
• OpenRouter: Terms of Service, including the model provider's terms it passes on, openrouter.ai/terms (Motion Cinema);
• BytePlus (ByteDance group): Acceptable Use Policy, which now includes its generative-AI rules, docs.byteplus.com/en/docs/legal/docs-acceptable-use-policy (Motion Cinema);
• MiniMax: Motion and Motion HD reach MiniMax's models only through fal, so fal's Acceptable Use Policy above governs them. MiniMax's API terms are not public; we have not found a separate MiniMax usage policy to link here, and will add it if MiniMax publishes one.

We have no partnership with and no endorsement from any of these companies. Their names belong to them and are used here only to tell you whose model runs your request.

We will update this statement when what we generate, or how we mark it, changes — and when we close the gaps named in sections 4 and 9. Questions about this statement: [email protected]. Reports of misuse: /report or [email protected]. Security issues: [email protected]. Who we are, and the current status of the operating entity, is at /imprint.