Speak in your own voice — in a language you don't speak

Here is the part of real-time translation that almost everyone gets wrong, and that almost no one talks about: the voice you hear.

You can have excellent speech recognition and excellent translation, and still end up with a meeting that feels like a machine reading a list. Because the last step — turning the translated text back into sound — is where most tools quietly substitute you with a single generic synthetic narrator. Eight people in the room, one robot voice for all of them. You lose who is speaking, the emphasis, the personality. Intelligible, but not a conversation.

InterMIND does the last step differently. When you speak, the other participants hear the translation in a voice that's recognizably yours — carrying your timbre and your way of speaking — now saying the words in their language. It isn't a flawless impression yet; the point is that it's you rather than a stock narrator, and it's getting better. This works for every participant, in both directions, at the same time.

This post is the missing chapter of Inside the four translation pipelines that run InterMIND: that piece explained how audio becomes translated audio. This one is about whose voice comes out the other end.

The default everyone ships, and why it's flat

If you've used live translation in any of the big meeting platforms, you know the sound. A neutral, evenly-paced voice reads the translation. It's the same voice whether the speaker is your CEO opening a town hall or a colleague cracking a joke. The technology underneath is text-to-speech with one fixed voice model, and the design assumption is that intelligibility is enough.

In a real meeting it isn't. Half of what a meeting communicates is who is saying it and how. Strip the voice and you've turned a discussion into a transcript that happens to be spoken aloud. People stop reacting to each other and start waiting their turn.

What InterMIND does instead

The translation runs as a cascaded pipeline — three specialized stages in sequence rather than one model trying to do everything. The first two stages are covered in the pipelines post; the voice step is the one this post is about:

1. ASR — speech recognition. Your words are transcribed in your own language, as you speak, on our own engine — the media server that carries the call (Mind API, OVH France) — so translation can start before the sentence ends.
2. MT — translation. The transcript is grouped into stable sentence fragments — clauses — so translation can begin before you've finished the sentence, and each fragment is translated progressively into the listener's language.
3. Zero-shot TTS — voice synthesis. Each translated fragment is spoken back out using a sample of your own voice, and streamed to the listener.

It's that third stage — ASR → MT → zero-shot TTS — that produces the effect. "Zero-shot" means the system doesn't need a pre-recorded enrollment or a training session for your voice. It models your voice from the audio of the meeting you're already in.

The warm-up: how it starts sounding like you so fast

There's a chicken-and-egg problem hiding in "use a sample of your own voice." At the very start of a call, the system hasn't heard enough of you yet to model your voice well.

InterMIND handles this with a progressive warm-up:

• For roughly the first 5–10 seconds, while it's still gathering enough of your speech, each translated fragment is synthesized using the audio fragment that matches what you just said in your source language. The voicing is anchored to your real, immediate speech.
• Once there's a long enough sample — that 5–10 second mark — the system locks onto it and uses it to voice everything afterward.

In practice you don't hear a switch flip. The translation sounds more like you as the conversation gets going — not a perfect double of your voice, but clearly yours rather than a machine's, and improving as the model hears more. The combination of progressive translation (clause by clause, not sentence by sentence) and progressive voicing is what keeps the whole thing under the latency budget while still sounding human.

Record your voice once, and skip the warm-up

The warm-up above is what happens by default, with nothing set up. Since September 2026 there is an optional second path for signed-in users: a saved voice sample in Settings → Your voice in translation.

Recording it is one gesture. Press Record my voice, wait for Speak now, and say the numbers from one to ten in any order. Each number lights up on the card as the speech engine hears it; when all ten are lit, the sample is checked and saved by itself. There is nothing to play back or confirm, and no countdown: the card keeps the stretch of your take that holds the numbers. A quiet room and a headset give the best result.

What the saved sample changes: from your next meeting, the synthesizer voices your translation with it from the first fragment, so the other side hears you as you from the first sentence instead of after the warm-up. Under the hood it is the same zero-shot synthesis with a better starting point; the live warm-up still refines the voice as the call goes on.

The take is gated before it is stored. It has to be 3 to 10 seconds of clear speech (the synthesizer's input window): too quiet, clipped, mostly silence or drowned in background noise, and the card tells you what to fix and asks for another take. Numbers were chosen as the phrase for a practical reason: they are short, recognizable in every one of the 23 languages, and the engine can confirm it heard all ten, which turns "did that recording work?" into a visible yes.

Two voice samples, two lifetimes

This is the part a security or legal team asks about immediately, so here it is plainly. There are two different samples, and they live differently.

The live sample used for the in-meeting warm-up is ephemeral. It exists only for the live conference session, in service of voicing the translation, and it is stored nowhere. The Mind API and SDK that power the real-time session retain no data; everything temporary dies when the conference session ends.

The saved sample from Settings is stored with your account, for one purpose: it is sent to the translation engine each time you join a meeting so your translated speech can be voiced with it, and for nothing else. It stays until you press Remove on the card, which returns you to the standard warm-up immediately, and it is deleted together with your account. Guests cannot record one; it is a signed-in feature by design.

It's worth being precise about what neither sample is: not one of InterMIND's recording features. Recording a meeting's video and audio is a separate, deliberate action you take on purpose, with its own controls. A voice sample is an input to the speech synthesizer: transient in the live case, yours to keep or delete in the saved one.

Why no one else ships this

It's not that voice cloning is a secret. It's that doing it live, per-participant, in both directions, under a one-second budget, across 23 languages, without storing anything you didn't ask it to is a different problem than cloning a voice offline for a podcast.

The big platforms optimize their translation for caption coverage and a single safe narrator voice — that's the cheap, robust default at scale. Keeping each speaker's own voice means the synthesis stage has to track every participant independently and stay inside the same latency budget the rest of the pipeline lives under. We built the voice engine ourselves, on our own infrastructure, which is what makes that trade-off ours to make. (More on why the engine is our own code: What one InterMIND meeting is built from.)

Where this is going: lip-sync

Keeping your voice is one half of a bigger goal. The other half is your face.

Right now you hear the other person in their own voice, but if you're on camera, their lips still move to the words they actually said — in a language you don't read. The next step is lip-sync: re-timing the speaker's mouth to the translated audio, so that on your screen they appear to be speaking your language.

Put the two together and the whole point of this work comes into focus. Two people who share no common language sit across a video call and see and hear each other as if each were a native speaker of the other's language — same voice, same face, no interpreter in the middle, no robot reading a script.

To be clear about status: voice is live today; lip-sync is on the roadmap, not shipped. We're calling out the destination because it's why the voice work matters — own-voice translation isn't the feature, it's the first half of "talk to anyone, in any language, as yourself."

Where to hear it

Own-voice translation is live today, across all 23 voice languages — the same languages listed in the docs. There's nothing to turn on separately: when translation is enabled in a meeting, participants automatically hear each other in their own voices. We'll be honest about where it stands: today the voice is already recognizably you, and the resemblance is something we're actively pushing closer. Go listen and judge for yourself.

• Try the demo — runs the live voice pipeline against your audio in any of the 23 languages.
• See the quality numbers — the same production pipeline, scored monthly against FLORES-200, with the full distribution published per language pair.
• Record your voice sample — signed in, one take of the numbers one to ten, and you are voiced as you from the first sentence.
• How it works, in the docs — the short version of this post.

A translated meeting should feel like the people who are actually in it talking to each other. Keeping your voice is how it gets there.