A multilingual room does not need a booth per language. It needs one clean audio feed from the desk, one link the audience opens on their own phones, and a way to fix a misheard name while the talk is still going.
A talk, a service, a lecture or a council meeting has one person speaking and an audience that does not all think in that language. The usual answers are expensive, partial or private.
The expensive answer is an interpreter for each language. Even with one per language, that is a booth, a set of receivers and a specialist for every language in the room, and the specialists do not say the same thing. Each interpreter makes their own choices under time pressure: which word to use for a term, what to compress, what to drop when the speaker speeds up. On an ordinary talk the differences hardly matter. On a legal hearing, a council vote or a technical briefing they do: a defined term, a dosage or a clause can reach the French listeners with one meaning and the Arabic listeners with another, and no one in the room is in a position to notice. The partial answer is a subtitle feed on a single screen, in only one second language. The private answer is fifty phones held up to the room: fifty slightly wrong transcripts, each phone hearing the echo and the coughing as clearly as the speaker.
The alternative is to transcribe the speaker once, from a clean source, translate that one transcript into whatever the audience needs, so every language starts from the same words, and let each listener choose on their own device. The recording is kept when the event ends. Here is how the host and the listener each set it up, and what it costs.
A live transcription can hear a microphone on any device, an RTMP feed from OBS or any other encoder that takes a server URL and stream key, a YouTube live, an internet radio station or any URL that serves a stream. For a stream, VocalStack pulls the audio itself; nothing runs on your side.
The source matters more than any setting. A room microphone, or a phone on a table, hears the speaker plus the room: reverberation off the walls, chairs, the audience. Speech recognition degrades sharply as the voice gets quieter relative to that noise, which is why those fifty phones disagree with each other. A feed from the mixing desk carries the speaker's own microphone with little of the room in it, so the transcript follows the speaker, not the room. If the venue has a desk, take the feed from it.
The spoken language is detected automatically, and the speaker may switch at any time; naming it in advance helps on noisy recordings and heavy accents. Speaker labels can be off, on, or When Needed, which turns them on once a second speaker is heard. Once audio is coming in, the audience needs somewhere to find it.
A broadcast is the link: one address and a QR code that listeners scan to open the live translation and pick their language. The link is made from the name you give the broadcast. Rename it later and the old link keeps working, so a QR code printed on a programme does not go stale.
A broadcast is public by default: anyone with the link can open it. Make it private and set a password, and listeners are asked for it before the transcript appears.
One broadcast can be fed by several live transcriptions, so a conference with two stages still shares one link. The link gets people in; what keeps them there is getting the names right.
Speech recognisers are trained on how people usually talk, so when a sound is ambiguous they reach for the common word. A product name, a surname, a village, a chemical: each has a more frequent near-homophone, and that is what comes out unless the recogniser is told the rare word is coming. That is what a script is for. Add a script or slides before you start, and the names and terms in them win over what was misheard. Those names and terms are where recognition usually falls down on a specialist talk. During the recording the Script tab follows along like a teleprompter, lighting the words spoken so far.
That is the whole setup. On the other side of the link, there is even less to do.
A listener needs a link or a QR code, and a phone. They open the link in a browser and see The broadcast hasn't started yet until the host presses record, then Live. The first line appears in a moment. They pick from the languages the host chose (up to 10 per transcription, out of 100 available) and read the transcript as the speaker talks. The host keeps speaking in their own language. A listener can change language at any time; the transcript so far is shown again in the new one.
If a listener prefers to listen, each translated line is read aloud. By default each speaker gets a distinct stock voice of matching gender, so a panel still sounds like a panel.
The three stages happen in order, and each one waits for the one before it. Recognition needs a stretch of speech before it can commit to the words, so the transcript trails the speaker by a fraction of a second. Translation needs a whole phrase, because word order differs between languages and the end of a sentence can change the beginning, so it follows a second or two later. Speech needs the translated sentence before it can start, so a spoken line comes last. None of this is a fault to tune away; it is what any system doing these three things must do, including a human interpreter, who also waits for the verb.
With the speaker's permission, a host can clone the speaker's voice from thirty seconds of that person speaking alone (an MP3, M4A or WAV; a few minutes is better). The translation is then read in the speaker's own voice in every language. Treat a voice as the biometric data it is. Under the GDPR, voice data processed to recognise a person generally falls within the special categories of Article 9, and some US states, Illinois among them, regulate voiceprints separately. Get the speaker's explicit consent before the sample is recorded, not after; the voice can be deleted at any time. Cloning also enrols the speaker, so they are labelled by name in future recordings. A host can assign a voice to each speaker, or mute one.
A listener who arrives late sees everything said before they came in, then follows live. If the connection drops, the page catches up on its own once the connection returns. When the host pauses, listeners see Paused. The transcript continues when the host resumes. When the host stops, the page shows the recording's date and length, and the broadcast lists its past recordings for replay.
A listener can only read what the engine heard, and live recognition gets some of it wrong.
The host screen has a Fix tab for exactly that. Mark uncertain words colours and underlines the words the engine was unsure of; Click to correct opens a line for review. The tab walks through uncertain segments one at a time, so a second person at the desk can work through them while the speaker carries on.
AI-assisted auto-fix corrects likely mistakes as lines arrive, which helps with noisy audio or heavy accents. Every fix is marked and can be undone. When a host rewrites a word or a line, even one still being spoken, every listener sees the correction. A line whose confidence is too low is hidden from listeners as Audio unclear rather than shown wrong; the host still sees every word. A blank is a known gap; a confident wrong word is a rumour.
Those corrections do not vanish when the talk ends.
Everything is kept. The recording, its transcript with corrections and translations, and the audio land in the host's library. Replay plays it back exactly as it arrived, line by line, with its audio. Once the words are final, a title and summary are written in every language, and the host screen's Summary and Export tabs carry them straight out of the event.
Exports are SRT subtitles, PDF, Excel or JSON, with corrections and translations included, and the audio downloads as MP3. Every transcript is searchable in any language.
The Audience view shows who listened, in which language, and when they joined: often the first real count an organiser gets of who needed which language. Until then, the usual estimate was whoever raised a hand when asked, in English, whether they needed translation. Listener identifiers are deleted with the recording.
Speech recognition, speaker recognition and voice synthesis run on servers we operate, and your audio is not sent to any third party for processing. Translation, summaries and AI-assisted corrections work on the transcript text, not the audio, and that text is processed by a language-model provider. We do not use your audio, transcripts or voices to train models.
Retention follows the plan: on Pay as you go, source audio and voices are kept for 7 days and transcriptions for 30 days; Premium and Enterprise keep them for the life of the account. The other thing a plan decides is the bill.
Two numbers matter. On Pay as you go and Premium, live transcription is $0.50 per hour of audio, billed per minute, in the Dashboard and through the API alike; Enterprise has volume rates. Listeners and languages are free: a one-hour event read in five languages by two hundred people costs $0.50. Files are cheaper ($0.15 per hour) because a live transcription runs in real time for the whole event. Plans differ in concurrency, maximum recording length and start-up speed; on Pay as you go a cold start can take up to 15 seconds, so press record a little before the speaker does, roughly one "Can everyone at the back hear me?" ahead. The full table is on the pricing page.
The home page has a two-phone demo, and a new account in the Dashboard comes with a $5 signup credit and needs no card. At the live rate, that credit covers ten hours of broadcasting.
Deciding between this and the live translation built into everyone's phone? Read why a phone is not enough for a room. For a full event, the multilingual conference checklist covers the audio path, several stages on one link and the budget. Developers can do all of this from code through the API reference, at the same rate.
