Made by passionate people, for people

albadr.aiBlog
← All posts

Translator12 min read

Half a second behind the voice

Translator is our MVP demo of live AI speech translation between five languages. On a desktop processor from 2012, with no graphics processor, an English word is on screen 573 milliseconds after it is spoken, on average, and about one recognized word in seven is wrong.

In 2025 IWSLT, the research conference where live speech translation systems are evaluated, timed a professional interpreter with the stopwatch it uses for machines: one person, working alone, from the audio only. Interpreting Czech into English, the professional came out 4.16 seconds behind the speaker by that metric, against 1.51 seconds for a research system on the same material. The human judges still rated the professional higher.

Translator, our demo, puts a spoken English word on the screen 573 milliseconds after it is spoken, and its French translation covers that word after 607 milliseconds. Those are means over 99 recorded sentences, on a desktop processor from 2012 (an Intel Core i7-3770, four cores, 8 GB of memory) with no graphics processor in use.

An interpreter produces speech and copes with real conversation. Our figures are for text on a screen, measured on recorded sentences that were read aloud, and about one recognized word in seven is wrong. Translator is an MVP demo, public since 15 November 2026 at translate.albadr.ai, with quotas.

What Translator is

Translator is a web page. A person picks the language they speak and the one they want to read, presses Start and talks, and their words and the translation appear as text while they are still talking. It works in any of the 20 directions between English, French, Russian, Arabic and Chinese, five of the six official languages of the United Nations (the sixth is Spanish). The output is text on a screen, not a synthetic voice.

All of the work is done by AI: neural speech recognition and neural translation, several models at once for a language pair, in real time. None of them needs a graphics processor. The same models run on an ordinary processor in a small server, and inside a browser tab on the visitor’s own machine. Under its name the page says “Fully AI solution from albadr.ai”, and one of the cards on its first screen spells out what that means: “No person listens, types or translates. Like any AI, it can get a word wrong.” We can put a number on that last sentence. About one recognized word in seven is wrong on read sentences from many speakers, more with strong accents, noise or two people talking at once.

As a demo it is rationed. A network address gets four hours a day with the engines on the visitor’s own device and twenty minutes a day through our server, and a session ends after 90 seconds without a voice. There are no accounts, no payment, no price and no interface for a client’s systems yet.

Where an interpreter is missing

A House of Lords committee reported 8,958 unfulfilled requests for interpreters in the courts of England and Wales between January and September 2024. Where an interpreter can be found, the day has a price. United States federal courts pay a federally certified contract interpreter $617 for a full day, and for simultaneous work interpreters usually work in teams of two and relieve each other every 20 to 30 minutes, says the Swiss region of AIIC, their international association.

We quote the courts because they keep count. They are the last place for a demo with that error rate. The need we have in mind is lower in stakes and, we think, far more common. In the 2021 census of England and Wales, 880,000 residents could not speak English well and 161,000 could not speak it at all, whatever their language. Some of them explain what they came for at a front desk, or sit in a clinic’s reception on a day when no interpreter was booked. It is the same at a lecture, a town-hall meeting, a guided visit, a sermon or on a support line: one person speaks, others must follow, and nobody was going to pay for a team of two interpreters. These are our proposals. Nothing is deployed in any of them.

On how much delay people tolerate we can only quote others. People in conversation answer fast: in recordings of informal talk in ten languages, the answer to a yes-or-no question began 208 milliseconds on average after the question ended, and a pause of 600 milliseconds or longer before a reply makes listeners expect an unwelcome answer, according to a review of research on turn-taking. For a translated voice, Google’s Meet team has said that earlier chains of transcription, translation and speech synthesis were often 10 to 20 seconds late, and that “two to three seconds was sort of a sweet spot”. That is a product manager quoted on the company’s blog, about a voice, with no method given. We have not studied readers of our own page, and we claim nothing about comprehension at half a second.

One design that runs in two places

The page captures the microphone and draws the text. It and the engine files are plain files, served without any program running. A small program at the network edge answers the page’s questions and admits sessions, and a service keeps the usage budget. The engines, the neural models that recognize speech and translate it, are the only heavy part, and they can run in two places.

One is the visitor’s own device. The engines are downloaded once (218 MB for English to French, 213 to 451 MB for the other pairs) and run in the browser. In this mode the audio does not leave the browser at all, which matters because a voice says more than its words: the French data protection authority CNIL lists age, gender, health or emotional state and identity among what it reveals. What does travel to our service is text and numbers, the recognized words and their translation among them. The service handles them in memory and stores none of it. What it keeps is a ledger of the time used and notes on its own faults, and its logging is switched off. This mode needs the download and is only as fast as the machine. Each device times itself before the large download and is told if it would not keep up, and an older tablet browser that cannot run the engines is sent to the server.

The other is a small container with 2 processors and 3.5 GiB of memory. Audio is streamed to it and exists only in memory. No audio is written to disk or logged. The container sleeps when nobody uses it, and the first visitor after a sleep pays for that: on the live site, a start from sleep opens its connection in about a second and has its models loaded after seven. A visitor who was promised that the microphone stays on the device is asked before any audio goes to the server.

Doing without a graphics processor is what lets the server fit in a container of that size and the device side in a browser tab. The cost is that models small enough for a processor hear like small models.

The page also reports on itself. Every message of live text carries the moment its newest word was spoken, on the page’s own clock, so the delay it shows is the delay that user experienced, transport included. Text that may still change is drawn lighter with a dotted underline. The choice was between showing words early and never taking a word back. We show early and mark every change.

The whole public demo is designed to stay inside one hosting plan of 5 USD a month, under a budget the service enforces. A visitor is counted under a name that is a salted digest of the network address. When a share is used up the page says so, and says when it is worth coming back. Turning a visitor away until a stated time seemed better than an unplanned bill.

What we measured

The figures in the table come from the test part of FLEURS, a public set of read sentences that no setting was chosen on. The English column is a single run, of 11 November. The French column is the last of six runs made from 11 to 14 November while the French models were still being changed, and its word errors read between 13.3% and 15.4% along the way, so it is not as clean a held-out figure as the English one. All were played into the real system at the speed of speech, on the 2012 desktop processor, with the engines on the same machine and no network in between.

English to FrenchFrench to English
Recordings99100
A spoken word is on screen after, mean (95th percentile)573 ms (715)532 ms (683)
That word is no longer marked provisional, mean (95th percentile)836 ms (1126)791 ms (1085)
Its translation covers it after, mean (95th percentile)607 ms (755)574 ms (740)
Whole translation on screen after the voice stops, median (95th percentile)470 ms (684)410 ms (661)
Word errors in the text as it finally stands14.10%13.63%
Translation as it finally stands, BLEU and chrF36.18 and 63.3931.86 and 61.15
Settled translation words replaced when a sentence is declared final0.88%0.33%

Read the second row with the third. “On screen” includes text still marked provisional, and the third row says when the mark is gone. Computing is the small part: one step of recognition and one translation together take under 0.1 s on average on that processor. Most of the half second is the recognizer needing to hear what follows a word before it can write it.

The error and translation figures are for the text as it finally stands, after any revision. In French a revision arrives 888 ms on average after the voice stops (95th percentile 1,571 ms), and until then about one French word in five is wrong. On 115 Arabic recordings that no setting was chosen on, the text as it finally stands on the server has 6.9% word errors. The last row counts only replacements when a sentence is declared final. While a sentence is still open, 16 to 20% of the settled translation words change, and the page marks each change.

Seven more directions were measured only on development recordings, the ones our settings were tuned on, 19 to 60 per direction. In those runs a word is on screen after 539 to 612 ms. With a simulated network round trip of 60 ms, the English to French figure was 669 ms through the server (42 development recordings).

Device mode is slower on the same machine: with the engines’ browser builds standing in for a device and the same simulated network, 644 to 840 ms across the 20 directions (17 to 20 development recordings each). For English, accuracy held in a browser. In a proof on one PC and one browser, the English recognizer made 11.64% word errors in a tab against 11.48% natively on 60 held-out recordings (46 of the 60 transcripts were identical), and the translation model, given the same 60 sentences, returned the server’s translations. Arabic is different: on the device stand-in (17 to 20 development recordings a direction) it makes 11.6 to 13.1% word errors against 5.9 to 7.2% on the server, until the visitor switches on an option that fetches 174 MB more.

We did not measure spontaneous speech, noise, strong accents, two people talking at once, phones, real networks, or the speed of the hosted container. Russian, Arabic and Chinese have no held-out delay figures. No person has rated the translations.

What the industry ships and charges

Large vendors ship live speech translation today, each in more languages than our five. Google Meet translates speech between any combination of 63 languages on eligible plans, and its help page says translations “will be a few seconds delayed”. Apple’s Live Translation with AirPods does all processing on the iPhone, needs an iPhone 15 Pro or later, and has neither Russian nor Arabic on its list. Microsoft’s Interpreter in Teams sends meeting audio to the cloud. Neither Apple’s nor Microsoft’s page gives a delay in seconds.

Research publishes more. The IWSLT evaluation has a low-latency regime of 0 to 2 seconds and a high one of 2 to 4, by a metric that leaves computing time out, and expects submitted systems to run on a single NVIDIA H100 graphics processor with 80 GB. Its 2026 findings put computing at about 0.3 seconds more where that could be measured. None of this is like for like. Those systems are tested on other material and by other metrics. We draw no league table. All we claim is that a delay of this size did not require a graphics processor.

On 17 November 2026 Microsoft’s Azure listed real-time speech translation at $2.50 per audio hour. That is a price to a customer, and we have none to set beside it. What we have is a hosting cost. By our own simulation, not from traffic, the 5 USD plan carries about 470 hours of sessions a month on visitors’ devices and 10.3 hours on the server. Nobody had visited when those figures were computed, and we have seen no invoice. The device figure is large because the computing happens on the visitor’s processor: a device session costs the host only a few short messages a second while someone speaks, and none in silence.

How a client would integrate it, and where it has run

Today the demo is closed on purpose. The page cannot be embedded in another site, and calls from other sites are refused. On request we can offer a client the connection underneath. It is one streaming connection with a short vocabulary: audio frames go up, and events come down (ready, live text, final text, revised text). A client’s display has to do with those events what our page does: show live text as text that may still change, then the final text, then a revision if one arrives. The server side runs as a single process or as a container image, on a processor and nothing else. Device mode is there for clients who want the audio never to leave the user’s machine. Invitation links carry their own minutes, outside the public quotas.

The rest would be built with the client, and we list it as work because none of it exists: keys and per-client metering, permission to embed, connectors to the client’s meeting, telephony or event systems, more languages, and output beyond text on a page.

On 11 November 2026 we took the first measurements, with the whole product as one process on a desktop PC. A day later the same engines were proven in a browser tab, and the edge version ran locally. By 13 November all 20 directions ran with the engines’ browser builds on that PC, six of them in a real browser tab. The public launch followed on 15 November, on an edge platform plus a container, with all five languages on the server the same evening. So between 11 and 15 November it ran as one process on a PC, in a browser tab, in an edge runtime and in a Linux container. A new server has to offer a processor and room for one process or one container. We have no recorded run on another cloud, and no figures from phones. Changes are checked against seven reference traces (5,496 events) that must come out the same byte for byte, beside more than 270 automated tests.

What the demo cannot do yet

Translator is an MVP demo. About one recognized word in seven is wrong on read sentences, and we have no figures for spontaneous speech, noise, accents or phones, nor for delay over a real network. It was tested with one speaker and one microphone, and overlapping speakers are not handled. There are five languages, quotas and no price. Device mode begins with a download of 213 to 451 MB, and an older tablet browser falls back to the server. The interface exists in all five languages, in translations that native speakers have not yet read. The capacity figures come from a simulation.

The demo is at translate.albadr.ai. Speak a sentence, then open the settings panel and read the delay the page measured for you. If you have a room where one person speaks and others must follow, or a system this should connect to, write to contact@albadr.ai.