--- title: "CAIO · AI video dubbing | AimRank" url: https://www.aimrank.io/caio/ description: "CAIO dubs your video into ten European languages in your own voice, lip syncs it, captions it, publishes it and measures the reach each language earns. Built for the languages dubbing tools get wrong. Access by request." language: en updated: 2026-10-21 --- CAIO · Video translation and dubbing # Built for the languages dubbing tools get wrong CAIO dubs your video into ten languages in your own voice, lip syncs it, captions it, publishes it and measures the reach each language earns. We run it for you. Most dubbing tools are tuned on English, Spanish and German, then pointed at everything else. We went the other way. Lithuanian has seven grammatical cases, Latvian seven, Estonian fourteen, and translated Baltic runs longer than its English source, so every shortcut in a dubbing pipeline breaks there first. Fix it there and German, Spanish, French, Italian, Portuguese and Polish are the easy direction. Already a customer? [Sign in ↗](https://caiociao.com) No commitment · reply within 48h · invite only, we set it up with you **10** languages, as source and as target **6** target languages per job **10 min** per video, while CAIO is invite only **3** lip sync levels, or none for voice only **Why there is no signup button** CAIO has no public sign up. Accounts are created by us, which is why the app tells everyone else to contact their administrator. Invite only is not a queue you wait in. It is the setup call you get. ## What comes back One upload. Up to six languages per job, rendered at the same time rather than one after another. ### A finished video per language An MP4 for every language you picked: your speaker's voice cloned into that language, the shots with a face on camera lip synced, and the original music and room sound separated out and mixed back under the new speech instead of thrown away. ### Audio only, or a YouTube pack Podcasts and voice work come back as an audio track per language. For YouTube there is a Multi Language Audio package: one video file plus one dubbed audio track per language, zipped the way YouTube Studio wants it. That package swaps the audio only, so it carries no lip sync. ### Captions and a thumbnail that reads Burned in captions in eight styles and eight fonts, with position and colour controls. Upload your cover image and the text on it is translated for each language. Every language also comes with SRT and VTT subtitle files. ### Shorts cut from the long video Up to five vertical clips per long video. A model scores candidates on hook, whether the clip stands on its own, emotional peak, quotability and pacing. Cuts snap to the silence between words so nothing lands mid syllable, the frame is reframed to 1080x1920 with the speaker tracked, and captions are burned in. A clip can then be dubbed like any other video. ### Slides and carousels Text inside carousel images is read, translated against your glossary with numbers and names left intact, and rendered back into the image. Up to ten slides. ### A provenance record A C2PA manifest on the dubbed video naming the models that touched it, the job and the target language, readable inside the app on every finished dubbed video job. It ships today as an unsigned sidecar file. More on that below. **And how fast?** Every language renders in parallel, so six languages is not six times one video. We do not publish a turnaround number, because the honest one moves with your video length, how many languages you asked for and the lip sync level you choose. Ask about a real video and you get a real answer for that video, not an average. ## How we run it Invite only means somebody sets it up. That is us. 01 ### You request access Tell us what you publish, where you want it to land and in which languages. We reply within 48h. If it is not a fit, we say so. 02 ### We set your brand up with you Your glossary: terms pinned to one translation, brand names on a never translate list, and the declined forms a name needs in Lithuanian, Latvian and Estonian. Your tone per market, seeded for all ten languages and overridden for your brand. Your voice, either recorded in the app or lifted from your own footage. Your accounts, connected. 03 ### You send video Up to 10 minutes per video while CAIO is invite only, and up to six target languages per job. MP4, MOV or WebM for video. MP3, WAV or M4A when the source is audio. 04 ### We render it Transcribe with word level timings. Look at the video. Translate against your glossary. Fit the timing per language. Speak it in the cloned voice. Lip sync the shots that need it. Mix the background back in. Burn the captions. Localise the thumbnail. Mark the file. Every language at once, and every stage is a separate step you can point at when something looks wrong. 05 ### You publish, we both read the numbers Schedule or post now, then watch reach and engagement per language next to your source language. If a market is flat, that is a glossary and market tone conversation, not a mystery. ## Why the output holds up Five mechanisms, not five adjectives. Where something was measured, the method and the date travel with the number. ### The translator watches the video Before a word is translated, keyframes from your video go to a vision model and what is on screen is written into the translation prompt. It settles what a transcript cannot: in Lithuanian the hazel tree and the hazelnut in the speaker's hand are different words, and only the picture knows which one is on screen. Live in production since 17 July 2026, on by default, and the scene note is cached so a six language job pays for it once. ### Names have to bend Lithuanian has seven grammatical cases, Latvian seven, Estonian fourteen. A naive dub leaves your brand name in its dictionary form every time, and that is the giveaway. Real Lithuanian says į AimRanką and AimRanke, not AimRank six times in a row. Your glossary can hold the declined forms of a term, so the translator picks the form the sentence needs instead of repeating the nominative. ### Baltic sentences run long Our timing budgets are set on the assumption that translated Baltic runs 15 to 20 percent longer than its English source. That is a working assumption from building for these languages, not a figure we have published a measurement for. Force every language into one timing band and you get needless retranslation and heavy time stretching that you can hear. So the budget is set per language: Lithuanian 18 percent, Latvian 16, Estonian 15, against a 10 percent default everywhere else. And the dub is never cut to fit the picture. The picture gives way instead. ### Facts are not localised Numbers, currencies and named real people stay exactly as recorded. 50,000 dollars does not quietly become 50,000 euros, and a named athlete does not become a more local one. Swapping a cultural reference happens only in the creative style, and you pick the style per job: literal, natural or creative. ### Lip sync is a choice, not a setting Three levels, picked per job, or off entirely for voice only dubbing. Each level is a different specialist model rather than a quality slider on one, so it is the biggest lever you have on both look and cost. Podcasts, voiceover and screen recordings do not need it at all. ## Measured, not asserted Every number here carries its method and its date. Where a measurement is directional rather than tight, it says so. **80.6%** of brand terms held with a glossary, against 50 percent without one A/B against a no glossary control, 9 languages, pass^k at k=2, meaning a term had to survive two independent runs rather than one lucky one. Measured 28 June 2026\. k=2 is a directional reliability estimate, not a tight confidence interval. **100%** of brand names and never translate terms held, in the same run Same test, 28 June 2026\. Terms pinned to one specific translation held at about 100 percent in German, Spanish, French, Italian and Portuguese, and at 75 percent in Polish. **3 to 4** distinct correct declensions of a brand name produced in Lithuanian, Latvian and Estonian, chosen by grammatical role Same test, 28 June 2026\. Case form injection is the part we were least sure would survive contact with a live model. **60.02s** is what one 60 second source came back as, in every target language 10 videos from one Lithuanian creator, dubbed into 6 languages: 60 audio tracks, 0 failures, 18 June 2026\. The one roughly 60 second source ranged 60.02 to 60.04 seconds across languages. That run was audio only, with no lip sync, and both the speech and the translation models have been replaced since. ## Dubbing is the first half A dubbing tool hands you files. What usually kills the project is what comes next: somebody has to write ten posts, schedule ten posts, and find out months later whether any of it worked. That part is in the same product, not a second tool you have to wire up. ### Connect the accounts Instagram, Facebook, Threads, LinkedIn and YouTube. ### Write once, adapt per platform One draft, tailored per placement, media attached. A caption generator writes title, description and hashtags for eight placements: Instagram Reels and Feed, Facebook Feed and Reels, Threads, LinkedIn, TikTok and YouTube Shorts. ### Schedule it Put a post on the calendar and it publishes itself at the time you set, or post it straight away. ### Measured reach, not modelled Dubs, posts published, reach, views and engagement per language, plus total reach against the reach your source language earned alone, counted from the accounts you connected. No euro estimate, because we could not defend the assumptions behind one. ## Languages and limits, in full Ten languages, and they work in both directions: your source video can be in any of them, and so can the dub. LithuanianLatvianEstonianEnglishPolishGermanSpanishFrenchItalianPortuguese Nothing outside that list, in either direction. Ten is as far as the market tone, the declined forms, the timing budgets, the on screen notice and our own testing actually reach. Other vendors will tell you eighty. We would rather tell you ten and mean it. ### The limits, before you ask Target languages Ten, listed above. Up to six per job, rendered in parallel. A seventh is refused at submission, not quietly dropped later. Source language One of the same ten, and it cannot also be one of your target languages. Video in MP4, MOV or WebM. Up to 10 minutes per video while CAIO is invite only, checked before the job starts. Audio in MP3, WAV or M4A. Out One MP4 per language, an audio only track per language, or a YouTube Multi Language Audio package. Lip sync Three levels, or off. The YouTube package swaps audio only, so it carries none. Slides and clips Ten slides per carousel. Up to five vertical clips per long video, at 1080x1920. Accounts Instagram, Facebook, Threads, LinkedIn, YouTube. TikTok post text is written for you, but the TikTok connection is not live yet. Where it runs AWS Frankfurt. Translation runs on EU Bedrock inference endpoints. Other stages call specialist providers outside that scope, and we will tell you which. Read the honest list below. Access By request. No signup, no trial. Run to our own standard ## One product, two proofs The EU AI Act's Article 50 transparency obligations took effect on 2 August 2026, and AI dubbing sits squarely inside them. So we run CAIO the way we tell clients to run AI: compliance checked with iCOMPLY, quality tuned with AimRank evals. The assurance we sell, proven on our own product first. ### What that looks like inside the file Dubbed video comes back with a notice burned into the frame, in the language of the dub: AI dubbed, Dubliuota DI, KI-Synchron, Doblado con IA, Doublage IA. The badge is always on. There is no per job switch and no brand styled version, because it is a regulatory marker and not a design element. The dubbed video and audio we deliver also carry a machine readable disclosure tag in the file's own metadata, and the dubbed video carries a C2PA provenance record naming the models that touched it. **And the part we will not round up** That C2PA record ships today as an unsigned sidecar file. The signing path is built and is not switched on, and we will not call it signed until it is. Audio only deliverables and the YouTube package carry the machine readable tag but no C2PA record. Carousel and thumbnail images carry no marking yet, and that decision is still open. The badge itself is written on a pass that is deliberately allowed to fail rather than hold up your file, and we alarm on that in production instead of pretending it cannot happen. If a dubbing product tells you its provenance is tamper proof, ask which mode its signer is running in. [Explore iCOMPLY ↗](https://icomply.aimrank.io) [See the methodology →](https://www.aimrank.io/methodology/) ## Before you ask Read this before you request access. It is cheaper for both of us. - **There is no signup** No trial, no self serve. Accounts are created by us and set up with you. If invite only is a dealbreaker, this is the wrong product today. - **There is no price on this page** What we run for you depends on how much video you make, in how many languages, at what lip sync level. We would rather size it with you than publish a number that is wrong for you. Ask and you get a number for your volume on the first call. - **We will not quote you a turnaround** Wall clock time moves too much with length, level and language count for one figure to mean anything, and we are not going to invent one. Ask about a real video and you get a real answer for that video. - **There is no comparison against another product** We ran one in April 2026, it measured a pipeline we no longer run, and a stale comparison is not evidence. - **There is no sample reel here** The dubs we are proudest of are other people's faces and voices, and we do not publish those without a written release. Ask on the call and we will show you the work. - **We did not build the lip sync models** We route to specialist providers, tune around them and measure them. What we did build is the translation layer, the per language timing budgets, the glossary and the disclosure. - **We will not tell you your media never leaves the EU** The infrastructure runs in AWS Frankfurt and translation runs on EU Bedrock inference endpoints. Transcription, voice, lip sync and image work call specialist providers outside that scope. Ask and we will tell you which stage goes where. The scoped claim is the honest one and the blanket claim is not. - **Voice consent is not yet asked on every dub** Saving a voice to your library asks you to confirm that you have the right and the consent to clone it, and records that confirmation. A regular dub clones the voice of the speaker in your video and does not ask that question in the app yet, so we ask it during setup instead. In both cases it is a confirmation, not a verification we perform. - **Ten minutes is the cap, for now** While CAIO is invite only, the system takes videos up to ten minutes and checks that before a job starts. The pipeline is built for longer sources, but we have not yet run one longer than ten minutes in production, so the cap stays until we have. If your videos run long, say so when you request access and we will test yours before either of us commits to anything. - **None of this is a compliance certificate** We ship the controls and describe them precisely. Whether your use of them meets your own obligations is a question for you and your counsel. ## Tell us what you want to dub One video, the languages you care about, and what you are trying to reach. We reply within 48h, and if CAIO is the wrong tool for it, we will say so. No commitment · reply within 48h · invite only, we set it up with you