Real-time transcription and translation explained: how live speech to text works, how far behind it runs, what affects accuracy, and what it costs.
Real-time transcription turns speech into text while the person is still talking. Words appear on screen moments after they are spoken, settle into final lines, and can be read by anyone following along. Add real-time translation and each reader can follow the same speech in their own language, as text or read aloud.
This guide covers real-time transcription and translation from both sides: what a listener sees, where the audio can come from, what decides accuracy, what developers get from an API, and what it costs.
Real-time transcription, also called live transcription or live speech to text, is speech recognition that runs while the audio is still arriving. A recorded file can be transcribed with the whole recording in hand. A live transcription has to commit to words with only the last few seconds to go on, and keep up with the speaker for as long as they talk.
Live captions are the same thing shown as short lines over a video or on a screen.
On VocalStack a live transcription can listen in 100 languages. It detects the spoken language by itself and follows the speaker if they switch, which how automatic language detection works explains, along with when to name the language in advance.
Text arrives in two stages. The newest words appear first as a draft, which can still change as more audio comes in. A moment later the draft settles into a final line that does not change again.
The draft exists because a recogniser decides a word partly by what follows it. "Ice cream" and "I scream" sound alike until the next word arrives. Showing the guess early is what makes a transcript feel live.
How far behind the speaker the text runs depends on the stage:
Two delays sit outside that chain. When the source is a public stream rather than a microphone, the stream itself already lags its live edge: a second or three for a plain internet radio stream, and anything from several seconds to half a minute for HLS. And on Pay as you go, a live transcription can take up to 15 seconds to start. The full breakdown, including the Auto, Fast and Accuracy latency modes, is in live translation latency, explained.
Translation runs on top of the transcript. A live transcription can be translated into up to 10 languages at once, chosen from 100. Every listener picks one of those on their own device and can change it at any time; the transcript so far is shown again in the new language.
Every language is translated from the same transcript, so every listener gets the same words, and a fix reaches all of them.
A listener who would rather listen than read can have each translated line read aloud. By default each speaker gets a distinct stock voice that matches their gender, so a panel still sounds like several people. With a speaker's consent, a host can clone that speaker's voice from about 30 seconds of them speaking alone, and the translation is then read in their own voice.
The rule for voices is narrow on purpose. A voice only ever reads what that same person said, translated into a language the listener picked. It never reads text someone types in, and it never reads one speaker's words in another speaker's voice.
How this works for a whole room, from the host's setup to a late joiner, is in one link, every language.
A live transcription needs an audio source. These are the ones VocalStack takes today:
Captions inside YouTube's own player, through its CC button, are coming soon. Until then, viewers of a YouTube stream follow the captions at the broadcast link.
If you need live transcription on Windows today, the Chrome or Edge extension, the OBS plugin and the Dashboard in a browser all work there.
A broadcast is how other people follow a live transcription. It is one link and a QR code. Listeners open it in a browser, pick a language and read along. They need no account and install nothing, and they cost nothing: the bill does not change with the number of listeners or languages.
A broadcast is public by default and can be protected with a password. If you rename its link, the old one keeps working, so a QR code printed on a programme does not go stale. One broadcast can carry several live transcriptions, so two stages can share one link.
A listener who arrives late sees everything said before they came in, then follows live. Afterwards the Audience view shows who listened, in which language, and when they joined.
For events, the multilingual conference checklist covers audio, stages and budget, and live transcription for churches, lecture halls and council chambers covers the weekly version of the same setup.
Everything above can be done from code. For live audio in your own app, the API has realtime sessions: your server makes one call to POST /v1/realtime/sessions with an API key and gets back a short-lived streamUrl. Your client opens a WebSocket to it, sends raw audio as it is captured, and reads the transcript back as JSON frames.
The frames follow the same two stages a reader sees. A transcript.draft replaces the live tail, a transcript.committed adds a final line, and with translation on, whole sentences arrive with their translations attached. Target languages, speaker labels and the latency mode can be changed mid-session with session.update.
A realtime session stores nothing. When you want the words kept, corrected and shared with an audience, you publish the audio to a live transcription and a broadcast instead. Official SDKs for TypeScript, Python and PHP wrap both, and an MCP server lets an AI assistant use the same account.
The WebSocket tutorial builds a working caption view step by step, and the API reference has every field. API use is billed at the same rate as the Dashboard.
Accuracy is decided mostly before any recogniser runs, by the audio it is given. Four things make the most difference.
The microphone. A microphone close to the speaker, or a feed from the venue's mixing desk, carries the voice with little of the room in it. A phone on a table hears the speaker plus echo, chairs and coughs. This is the biggest single factor, and why a phone is not enough for a room explains why.
Noise and overlap. Background noise causes dropped words and invented ones, and two people talking at once confuse both the words and who said them. Audio transcription errors lists seven common causes, what each looks like and how to fix it.
A reference script. Well-known names and terms are recognised as they are. A company's product names, an internal acronym or a guest's unusual surname cannot be known in advance, so add a script or slides before you start, and the names and terms in them are written exactly as you spelled them.
Speaker labels. Labels can be Off, On, or When Needed, which turns them on once a second speaker is heard. Speaker diarization vs speech recognition explains how the two answers, what was said and who said it, are separate and fixed separately.
The latency mode matters too: Fast gets text out sooner but revises more of it, Accuracy waits a little longer.
A single accuracy figure says less than it seems to. Word error rate counts a wrong name the same as a dropped "the", which is why what word error rate measures, and what it misses argues for judging a system by which errors it lets you catch and fix.
No live transcript is perfect, so the host screen has a Fix tab. It can mark the words the recogniser was unsure of, and any line can be opened and corrected. A second person at the desk can work through them while the speaker carries on.
When the host rewrites a word or a line, even one still being spoken, every listener sees the correction, in every language. A line heard with too little confidence is hidden from listeners as "Audio unclear" rather than shown wrong; the host still sees it.
The recording is kept with its transcript, corrections and translations. Afterwards you can:
How long audio and transcripts are kept depends on the plan, and the pricing page lists it.
Most live transcription tools look alike in a demo; the differences show up in the questions below. The right-hand column describes VocalStack as it is today.
| Question | Why it matters | VocalStack |
|---|---|---|
| Which languages can it hear? | A tool that cannot transcribe the speaker's language cannot help, however good its translation. | 100 languages, detected automatically, with switching mid-talk |
| How far behind the speaker is the text? | Readers notice a growing gap more than an occasional wrong word. | Original text a fraction of a second behind, translation a second or two later; three latency modes |
| Does it translate, and who picks the language? | One subtitle language on one screen leaves everyone else out. | Up to 10 languages per transcription, each listener picks theirs |
| Can it speak the translation? | Some listeners cannot read along, or would rather not. | Stock voices by default; a speaker's own cloned voice with their consent |
| How does the audience get it? | Apps and accounts lose people at the door. | One link and QR code, no account, optional password |
| Can you fix mistakes live? | A misheard name repeated for an hour does real damage. | Live corrections pushed to every listener |
| Where can the audio come from? | Your setup should not have to change to fit the tool. | Microphone, browser extensions, desktop app, OBS, RTMP, stream links |
| Is there an API? | You may want captions inside your own product. | REST and WebSocket API, SDKs, MCP server, same rate as the Dashboard |
| How is it priced? | Per-listener or per-seat pricing grows with your audience. | Per second of audio; listeners and languages free |
On VocalStack you pay per second of audio transcribed. At the time of writing, live transcription costs $0.60 an hour and files cost $0.20 an hour, on every plan, in the Dashboard and through the API alike. Translation, voices, summaries, broadcasting and listeners never add to the bill, so a one-hour talk read by 200 people in five languages costs the same $0.60 as one read by nobody.
The plans differ in a few ways:
If nobody needs to read along as it happens, upload the recording afterwards at the file rate. Current prices, retention and limits for each plan are on the pricing page.
A new VocalStack account comes with $5 of signup credit and needs no card. At $0.60 an hour that is about eight hours of live transcription, with translation and broadcasting included. After that you pay per second. Listeners are always free.
Yes. A live transcription can be translated into up to 10 languages at once while the person speaks. The translation appears a second or two after the original text, and each listener picks their own language.
It depends mostly on the audio. A close microphone or a desk feed, little background noise, one person speaking at a time and a reference script with the names in it give the best results. Live corrections let a host fix what is still wrong while the talk goes on.
The original-language text trails the speaker by a fraction of a second, and translations follow a second or two later. A public stream adds its own delay before the audio reaches VocalStack.
Yes, through the Chrome or Edge extension, the OBS Studio plugin, or the Dashboard in a browser. The desktop app is on macOS today, with Windows coming soon.
No. Listeners open the broadcast link or scan its QR code in a browser, pick a language and read. There is nothing to install.
Open the Dashboard, start a live transcription from your microphone and open its broadcast link on your phone. The signup credit covers your first hours, and you can see the drafts settle and the translation follow for yourself.