A person with a face, a voice, a mind and a pair of eyes. And it is yours.

How the service is right now

This line is written by the service, not by us: GET /salute is open and needs no key, and you can repeat it now. It also says what is missing.

The people already built

Press the video on a card and you hear her speak: no login needed to watch. You only need an account later, when you want to talk back.

Loading the people from the shop window…

The three faces are photorealistic invented portraits from a commercially licensed model: none of these people exists, and no real face was used to make them. Model, licence, seed and prompt are published in the repository (banco/dataset/persone/provenienza.json).

What it can do

• Speaks You give it a text, it gives you back a video of its face saying the words, with the audio and the lips in sync.
• Holds a spoken conversation You talk into the microphone: it transcribes what you said, thinks out an answer and hands it back as video, with subtitles.
• Sees Turn the camera on and send it a frame: it describes what it sees and takes it into account in the next answer.
• Is driven by API The same functions, from your own programs: curl, Python, JavaScript in the browser. The key is minted from the account area. Start in 5 minutes is the six real commands; the API documentation is the full contract.
• Becomes you From the recording room you film yourself in the browser for twenty seconds: it clones your face and your voice and saves the avatar in your account.
• Stays where you are All of it runs on our hardware in Europe. No outside service, no bridge to American suppliers.

Comparison with Tavus: the numbers we measured, and where we are behind

We speak of Tavus only through our own measurements, taken with the shared ruler on real files; we do not quote their prices, because prices change and are not ours to publish. The last five rows are the ones we lose: a comparison won on every row would convince nobody, and rightly so. Where it says bench the number is the drawing measured on its own, on one card and with no network; where it says real turn it is what a caller actually waited.

• First frame, the drawing on the bench: 310 ms How long the engine takes to draw the first frame that can be shown: one card, no network, no voice in front. It tells you what the engine is worth — not what a caller waits, which is further down.
• Speed of the drawing, on the bench: 3.27× real time One second of video costs less than a third of a second of card: it is the margin the cost per minute rests on. It is the margin of the drawing: the whole chain, on a real turn, does not hold it yet.
• Sharpness: 1.2259 This was the defect that kept us out of the race. We do not credit the GPU with the gain: a card moves milliseconds, not pixels.
• Minutes counted: 0.75 % apart What we bill you is checked against an independent ruler that looks at the delivered file, not against an estimate from the byte count.
• Head composure: 0.233 against 0.210 11 % above its only comparable witness: inside the band they publish, above the one clip we can honestly measure against.
• And that green is not robust: 0.028 against 0.007 The uncertainty of our own editing tool is four times the margin the verdict would rest on. We call that a draw, not a win.
• Short answers, out of band today: 0.377 on 2 seconds All of the excess sits in the first second, and 2-5 seconds is the normal length of a conversational answer. In steady state we match; coming out of the still pose we do not.
• First frame on a real turn: 1339-2383 ms This is the buyer's number: the wait from the POST, with the voice in front and the face on a rented pod. The target we set ourselves is p50 and p90 both under 500 ms in the browser — as percentiles, not as a ceiling: a promise that holds at the median and is lost one turn in ten is not a promise — and today we are two to four times above it. And it is two numbers: the instant the screen stops being empty, and the instant speech begins; the first one is pulled down by the mute head sent ahead, and on its own it would call kept what the caller does not have yet. The cause is measured and it is not the card: it is the instant at which the audio sampling rate becomes known. And our n is 2: we will publish a p90 above twenty real turns, not before. Until it is closed, it is written here.
• Speed on a real turn: 0.82× real time The drawing alone runs at 3.27×; the whole chain — voice, lips, muxing — on a real turn sits below real time, and a player consuming faster than the stream arrives would stall. Today we cover it with a living base that never freezes; the real margin is the wall we are working on.
• Cold start: 73.6 s Paid once per engine start, not per session. It must be closed before production, and until it is, it is written here.

Who it is for

• Organisations whose data cannot leave Healthcare, schools, public administration, professional practices. If the privacy notice you signed does not let you send a person's face to a supplier outside Europe, this is the way.
• Software vendors who want to resell it Access is by API key: your product calls ours, with your brand and your colours in front. Your customers never need to know we exist.
• Anyone looking for a second source Our calls follow the pattern this industry already uses — session, speak, converse, see — because changing supplier should cost a configuration, not a rewrite.

The plans

The price list is not published and the product is not yet purchasable online: the price of all three plans is on request. What is already true, and verifiable from your account, is the unit of measure: we count minutes of face delivered — not calls, not bytes — and the count is checked against the real duration of the file (0.75 % apart).

• Prova (trial) available Self-service account, no credit card. One API key with a daily quota, the people in the catalogue, and usage you read from the account in minutes of face.
  Price: on request.
• Standard under construction A licence per organisation, with several keys under it: permissions per key (read, speak, converse, cloning), a quota per key and per account, a usage log row by row, and isolation between organisations. The isolation is already proven; the licences that hold the keys under them are the piece we are writing.
  Price: on request.
• Enterprise planned The engine on your own hardware or on dedicated cards; a clone of your face and your voice, with consent on record; the versioned OpenAPI contract, keeping today's aliases; the licences of the weights declared row by row.
  Price: on request.

Ask for a licence

The price comes back by email, because the price list is not published. We keep only what you type here: the request stays on this service and goes to no outside supplier, and your IP address is not written anywhere.

How to start

• 1. You log inWith a username and a password, from Log in. If you have no account you create one from Sign up: an email and a name are enough.
• 2. You look at the galleryThe people that come with the product, each with a portrait and a short introduction video — watched up there, without logging in.
• 3. You build your avatarTwenty seconds of webcam, and the avatar stays saved in your account: you find it again next time.
• 4. You talkWith our people or with your own avatar, in video chat, out loud, with subtitles.
• 5. You take the keyIn the API section of the account: the key and ready-made examples for curl, Python and JavaScript.