إشعار الذكاء الاصطناعي والصوت

Beta version v0.1. If it changes, we will tell you in the app.

Talkalike uses artificial intelligence (AI) to help two people who speak different languages talk to each other. This notice explains, in short, where AI is used, which companies provide it, and how we tell you when something is made by AI.

---

1. Where AI is used

Feature	What the AI does	Provider
Text messages	Translates each message into the reader's language, using recent messages in the chat, your shared glossary and (if switched on) the chat memory as context	An AI model (currently DeepSeek V4.1 Flash) reached through OpenRouter and run by DeepInfra (United States), Nebius (Netherlands) or CoreWeave (United States); xAI (Grok, United States) as a backup
Chat memory (optional)	Writes a short summary of how you talk to each other, to improve translations. Both of you can see, edit, delete or switch it off. A voice message's words go into it only once their check found them clear.	The same as text messages
"About me"	Translates your profile text for the person who reads it	The same as text messages
Link previews	Translates the short description on the preview card of a link you send into the reader's language (not the link, the title or the rest of the chat)	The same as text messages
Voice messages	Turns the recording into text, translates it, and reads the translation aloud with an AI voice	ElevenLabs (speech to text and AI voice, United States); translation as for text messages
Live call translation	Turns what each person says into text as they speak (captions), translates each sentence, and speaks the translation with an AI voice	ElevenLabs (live speech to text), xAI (translation), Cartesia (AI voice), all in the United States
Explicit voice messages	Says whether a voice message's words are sexually explicit, so that an explicit one is locked for people whose age is not verified (see below)	The same as text messages
Explicit photos and videos	An image recognition model gives each picture a score, so that explicit ones are blurred for people whose age is not verified	Our own server. Pictures never go to a third party for this

Your choice. None of this happens to what you write, say or send unless you have given the optional consent described in our Privacy Notice (the images on link previews, which come from websites, are always checked). The explicit-content checks need that consent given to its current text, which names every kind checked; someone who gave it to an earlier text is asked again, and until they accept, what they send is translated but not checked. Without it, the other person sees your original words, with a short note that you turned off translation, and your photos, videos and voice messages are not checked, so they stay blurred or locked for people whose age is not verified. Messages you receive are still translated.

Explicit voice messages and pictures

Explicit photos, videos and voice messages are blurred or locked for people whose age has not been verified (see our Privacy Notice, section 6, and our Acceptable Use Policy).

• Voice messages. When a voice message is translated, the request also asks the AI model to say whether the words are sexually explicit (sexual acts, genitals or sexual nudity; flirting, love, swearing and hints do not count): the answer carries an adult field, true or false, next to the translation. It never changes the translation. When a voice message is not translated (for example because you both read the same language), we send one short extra request to the same providers with only the transcript and its language, and nothing else from the chat (no earlier messages, memory or glossary). Both happen only with the sender's consent. Without it, or if the answer is missing, not a clear true or false, or refused, the voice message counts as not checked and stays locked for people whose age is not verified. We count these requests for cost control, without their content. Written messages are never checked.
• Pictures. A model on our own server looks at photos, videos and other pictures. It is not a large language model and it runs only on our server; nothing goes to an AI company for it.
• What it is for. Both checks only decide whether something is blurred or locked for people whose age is not verified, and whether explicit content from a sender whose age is not verified waits until they verify. Nobody at Talkalike looks at the results. They can be wrong both ways, which is why you can also mark your own media as explicit and blur something you received.

2. AI voices are synthetic, and we label them

• The voice that reads a translated voice message or speaks a translation in a call is generated by AI. It is not the real voice of the person you are talking to.
• Today these are ready-made AI voices. We pick the one that best matches the speaker's "My voice" settings (such as gender, age and tone), or the voice you chose for that person. They are not copies of anyone's voice.
• Every AI voice in Talkalike is labelled "AI voice" where you play or hear it, and we explain this the first time you meet it. Translated messages show "Translated by AI" next to "Show original".
• You can always listen to the original voice message. In calls, you can still hear the other person's real voice quietly underneath the translation.
• Machine-readable marking. When Talkalike stores a translated voice message, it writes metadata into the audio file saying that it is AI-generated, which app and voice model made it, and when. If writing this mark fails for a technical reason, the translated voice is stored without it (it is not made again, and no warning is shown); it is still labelled "AI voice" in the app. In calls, the translated voice is sent on an audio track whose name starts with ai-voice, by a participant marked ai-voice. This is our best effort under Article 50 of the EU AI Act; we will look at stronger methods, such as watermarks, before a public launch.

3. Before a call is translated, we ask you both

Live translation means your speech is captured and processed while you talk. Before translation starts in a call, both people see a notice explaining that:

• speech is captured and turned into text by ElevenLabs;
• the text is translated by xAI;
• the translation is spoken by an AI voice from Cartesia;
• nothing from the call is recorded or saved by Talkalike.

ElevenLabs, which turns the speech into text, may keep logs of it under its own terms (see our Privacy Notice, section 10).

The first time, you see the full notice; on later calls, a short banner. You can continue with translation or call without it. Translation starts only when both of you have chosen to continue. If either of you calls without translation, or has not given the consent in our Privacy Notice, no one's speech is captured and the call goes on without translation.

We keep a record of each notice shown and each answer (who, which call, when and which version of the text) for 5 years, as ElevenLabs' speech-to-text terms require.

4. AI makes mistakes

AI translation and speech recognition can be wrong. They can miss nuance, mishear words, or occasionally change the meaning. That is why Talkalike always lets you:

• see the original of any message;
• see what the other person received ("Show what they got");
• ask for a new translation, and add words to your shared glossary;
• tap "Wrong translation" on a caption in a call to have that sentence translated again.

Do not rely on AI translation alone for medical, legal, financial or safety-critical matters.

5. What AI does not do

• AI does not make decisions about you or your account. Any decision about reported content is made by a person. The only automated effects are those of the explicit-content checks in section 1: media they find explicit is blurred or locked (people whose age is verified open it with one tap, others must verify first), and if they find a message's voice, photo, video, GIF, sticker or a picture or video sent as a file explicit while the sender's age is not verified, the message waits until the sender verifies their age, and is deleted for both people if they do not within 7 days. To avoid this, verify your age (once, with Yoti) before or within those 7 days. We do not think this has legal or similarly significant effects on you, and you can complain to us (see our Privacy Notice, section 6).
• We do not use AI to look for rule-breaking in your private chats. The explicit-content checks in section 1 are an age gate: their results are not used for reports, decisions about your account or anything else.
• We do not let our providers use your content to train their AI models. OpenRouter requests refuse hosts that train on data and ask for zero data retention; xAI zero data retention is on; ElevenLabs' use of our data is switched off; we have opted out of Cartesia's training.
• Translations are presented as translations. We never present AI output as if a person wrote or said it.

6. Voice cloning: only your own voice, only with your explicit consent (not available yet)

In future we may offer an optional feature that reads your messages in a voice like your own. If we do:

• it will be made only from a recording of your own voice that you make live in the app; you cannot upload a file or use anyone else's voice;
• it will be used only to read your own messages and call translations to the people you talk to; no one else can make your voice say anything;
• we will ask for your separate, explicit consent first, and tell you clearly that it covers all content you send, including sexual content;
• it will still be labelled as an AI voice (for example "AI voice · based on Sofía's voice");
• you can withdraw consent and delete the voice with one tap, and we will delete it from our provider's systems too;
• we will record when you consented and to which version of the text.

7. More information

• How your data is handled, and who receives it: Privacy Notice
• What content is allowed: Acceptable Use Policy
• Questions: our complaints form (Settings → About and legal → Complain)

---

Friendify Ltd · registered in England and Wales · company number 16656938 · 128 City Road, London, United Kingdom, EC1V 2NX