Live translation accuracy is mostly decided before any model runs, by where the microphone is. Take the audio from the mixing desk, not from a seat in row twelve, transcribe it once, and a mistake becomes one line someone can fix for everyone.
The translation apps built into phones are a real achievement. They often run a speech model small enough to fit on the device and its battery (some hand the audio to a server when there is a connection), and for short, clear, everyday sentences they do the job: asking for directions, reading a menu, a conversation across a counter.
A room full of people listening to one speaker is a different problem. The things that go wrong in real speech are exactly the things a small model is worst at: uncommon words and names, words that sound alike ("their" and "there", "sail" and "sale"), technical terms, accents, people talking over each other, and background noise. At a sailing club, a model that hears "sale" has made a reasonable guess and a confident mistake.
Two things make the small model guess more than a large one. It has been compressed to fit the phone, so it knows fewer words and fewer ways of saying them; it approximates a surname it has not met from the sounds alone. And live recognition has to commit to a word before the rest of the sentence arrives, so the model cannot use what would have settled "sail" against "sale": the words that come next, or a script that says what the talk is about. In speech recognition, context means exactly those two things: the surrounding words, and what is known in advance.
When the model is not sure, it guesses. The translation is then built on the guess, and the person reading it has no way to know. Nobody in the room can see that a line was wrong, let alone put it right. And a room is not one listener but fifty.
A phone hears the room, not the speaker. Distance is the reason: a lapel or desk microphone sits a few centimetres from the mouth, while a phone in row twelve is ten metres away or more, and most of what reaches it has bounced off the walls, the ceiling and the backs of chairs first. Reflected sound arrives late and smeared. On top of it the phone picks up the person coughing two rows back, the murmur of a neighbour, and whatever the loudspeakers do to the voice on its way across the hall. Every seat gets a slightly different recording, so every seat gets a slightly different transcript.
A phone also translates for one person. Fifty attendees holding up fifty phones produce fifty independent transcripts of the same talk, with fifty different sets of mistakes. Two people comparing notes afterwards will not agree on what was said. The speaker, meanwhile, is addressing a wall of raised phones, which is a hard crowd to read.
For some people in the room this is more than an inconvenience; it decides whether they can follow the talk at all. A hearing-impaired attendee, or someone following in a third language, needs one correct shared text, not fifty best efforts. That means transcribing the speaker once and giving everyone the same result. Transcribing once only pays off if it is done well, and that starts with where the audio comes from.
The transcript is only as good as the audio it is made from. So the first decision is not which model to use but which signal to feed it. The second is what to tell the model before it starts.
A live transcription takes its audio from wherever the clean signal already is:
VocalStack pulls a URL stream itself, so nothing runs on your side. An RTMP feed works the other way round: your encoder pushes it. The speaker's language is detected automatically, and they may switch languages at any time. If you know the language and the recording is noisy or heavily accented, naming it improves accuracy.
Clean audio removes the mistakes caused by the room. It does nothing for the mistakes caused by vocabulary, and those need a different kind of help.
A phone has no idea who is speaking or what the talk is about, so a product name or a surname is just another sound to approximate. A broadcast can be told what to expect in advance.
If you add a script or slides before you start, the names and terms in them win over what the engine misheard. During the recording the Script tab follows the speaker like a teleprompter. This is the single biggest accuracy gain for the kind of speech that fills conferences and lectures: dense with proper nouns and vocabulary that no general-purpose model has heard often enough.
Good audio and a glossary remove most of the errors. The question is what happens to the errors that remain.
Even with good audio and a script, some lines will come out wrong. The difference is what happens next.
On a phone, nothing. The wrong line scrolls past and the translation built on it goes with it.
In a broadcast, automatic help comes first. AI-assisted auto-fix, on by default, corrects likely mishearings as lines arrive, which helps most on noisy audio and heavy accents. A line too unclear to be worth showing is hidden from listeners as "Audio unclear" rather than passed off as fact.
Then there is the host, who has a Fix tab: the words the engine was unsure of can be coloured and underlined, a click opens a line for review, and any word or line can be rewritten while the transcription is still live. Every listener sees the correction in their own language. Every fix, automatic or human, is marked and can be undone.
The point is not that a server never makes mistakes. It is that a mistake is visible, owned by someone in the room, and fixed once for everyone.
Corrections fix what listeners read. Listeners who would rather hear the talk get the same corrected text, spoken.
A listener who would rather listen than read can have each translated line read aloud. By default each speaker gets a distinct stock voice of matching gender, so a panel still sounds like a panel. You can assign a voice to each speaker, or mute one.
With the speaker's permission, you can go further and clone their voice from thirty seconds of their speech, either uploaded as a clip or taken straight from a speaker in a recording; a few minutes gives a better result. The translation is then read in the speaker's own voice in every language. A speaker with a saved voice can also be recognised in later recordings, although the system would rather leave a speaker unnamed than name the wrong person.
Two rules apply, and they are deliberate. A cloned voice is biometric data: it is only made with the speaker's explicit agreement and can be deleted at any time. And a voice only ever speaks a transcript or its translation; it cannot be made to say something its owner never said.
Everything so far happens while the speaker is talking. The last difference shows up after they stop.
The fifty phone transcripts evaporate when the apps close. A broadcast keeps the recording in your library: the audio, and the transcript with every correction and translation. There it can be replayed line by line, searched in any language and exported; the walkthrough in One link, every language covers what the library does. Listeners cost nothing, however many there are. You pay for the audio transcribed, at $0.50 per hour for live, and the pricing page has the rate card.
A phone in every hand is still a fine fallback for a hallway conversation. For a room that has to understand one speaker together, the speaker should be transcribed once, from the source, with someone able to fix what goes wrong. You can see what that looks like in the interactive demo on the home page.
