How accurate this is, and when not to trust it.

1The short version

Everything is produced by machine while your event runs, and nobody checks it before your audience sees it. Usually it is good enough for an audience to follow along. It is not a guarantee.

• Names, numbers, specialist words, idiom and humour are where it slips most.
• Clean sound from a mixing desk, one person speaking at a time and a speaker close to the microphone make the biggest difference. Giving the venue your names and terms beforehand helps next.
• Do not rely on it where a misunderstanding cannot be undone, such as courts, medicine or safety instruction. Use a qualified human interpreter there.
• The audio is not recorded. The text is kept for three, six or twelve months, as you choose, and is never used to train or evaluate a model.

2Everything here is machine-made

Speech is recognised automatically, translated automatically and, where a venue turns it on, spoken aloud automatically. There is no interpreter in the loop and no editor between the microphone and the screen. What is produced goes straight to the room.

This is not a human interpreting service and it is not a certified or sworn translation. It produces a good working understanding of what is being said, very quickly, in many languages at once — which is a different thing from producing a guaranteed record.

3What that means in practice

Used well, in a room with decent sound and one person speaking at a time, the result is usually good enough that an audience follows along without having to work at it. That is what we aim for and what we most often see. It is not a measured guarantee, and we are not going to pretend we have one.

It will still make mistakes, and these are the ones that recur:

• Names. People’s names, place names and organisation names are where automatic systems are weakest, especially when they are not from the language being spoken.
• Numbers, dates and amounts. Usually right, occasionally not, and a wrong number reads just as confidently as a right one.
• Words with more than one meaning. The system picks a reading from the surrounding sentence, and in a short sentence there may not be enough around it.
• Specialist vocabulary. Technical, legal, medical or in-house terms the system has no reason to know.
• Humour, idiom and deliberate ambiguity. These often arrive flattened into their literal meaning.
• Register. A translation can land more or less formal, or more or less blunt, than the original was.

A mistake often affects a single line, and the next one is unaffected. Not always: poor audio, the wrong language being detected, or a name it keeps hearing the same wrong way can run through more than one line before it settles.

4Audio quality decides almost everything

In our experience the biggest single factor in how good the result is — ahead of anything else on this page — is what the service is given to listen to. A clean feed from your sound desk produces a noticeably better result than a laptop microphone at the back of a hall.

What helps most:

• Take audio from the mixing desk rather than from a microphone in the room. The mixer to computer guide shows how. In a small room with one speaker, a headset or USB microphone close to the speaker also works: see the single microphone guide.
• One person speaking at a time. People talking over each other is the hardest case for any automatic system.
• A speaker close to their microphone, at a steady distance, in a room that is not fighting them. The audio level guide explains how to check it.
• Music, applause and crowd noise between sentences are usually handled without trouble, though they can throw up a stray line. Music underneath continuous speech is a real problem.
• Steady pace and complete sentences. Trailing off mid-thought gives the system nothing to finish.

5What you can do before the event

Each venue can be given the words you most need it to get right — the names of the people who will speak, place names, your organisation’s own vocabulary and anything else particular to you. This makes a real difference to exactly the category that fails most often, and it takes a few minutes.

A venue can also list words that should never appear on screen, and choose whether they are hidden or removed. Automatic translation occasionally produces coarser language than the original warranted, and for some rooms that matters a great deal.

Where more than one language will be spoken, say so in the venue’s settings rather than leaving the service to work it out. Telling it to expect two languages produces a better result than restricting it to one and being surprised.

Run a session for a few minutes before the day with whoever will speak. It costs very little and it tells you what your room actually produces.

6Reading and listening are not equally reliable

Captions and spoken translation carry the same text and therefore the same errors. The difference is that a reader can see a line that looks wrong, stop on it and read it again, where a listener gets one pass and no way back.

Spoken translation adds a second automatic step on top of the first: the translated text is read aloud by a synthetic voice. It inherits every error in the text and adds the possibility of a word being pronounced oddly. It is genuinely useful — for people who cannot read a screen comfortably, or who would rather listen — and it is the output with the most between it and the original words.

Every language on the list has a voice. The voices were built from real speakers and carry an accent, so a native listener may notice one; what they hear is the translated text read in their language.

7When not to use this

There are events where automatic translation is the wrong tool, and we would rather say so than sell you something unsuitable. Use a qualified human interpreter where being misunderstood has a consequence that cannot be undone. In particular:

• legal proceedings, hearings, sworn evidence, police interviews, or anything a court will later rely on;
• medical consultations, diagnosis, consent to treatment, or anything clinical;
• safety-critical instruction — emergency procedures, evacuation, operating machinery, site inductions where getting it wrong injures somebody;
• contracts, financial advice, or anything where a person acts on the words and loses money if they are wrong;
• any situation where a law or a regulator requires a certified, sworn or accredited interpreter. This service is none of those.

It is also not a substitute for captioning provided as an accessibility accommodation where a standard or a statute specifies what must be provided. Check what you are required to supply before assuming this satisfies it.

The same caution applies after the event, not only during it. A transcript can be exported and read months later by somebody who was not in the room and has no idea which lines were wrong. It is a useful record of roughly what was said. It is not a verified one, and it should not be treated as the authority on what somebody actually said or agreed to.

Using it alongside a human interpreter is reasonable — for the languages the interpreter does not cover, or so the room can read as well as listen. Be clear with yourself about who is covered by which, though: anyone relying on a language the interpreter does not speak is relying on automatic output alone, and everything on this page applies to them in full.

8Tell the room what it is

We recommend telling your audience, in the invitation or on the screen before you start, that the translation is produced automatically and may contain errors. People are far more forgiving of a mistake they were expecting, and it is the honest thing to do.

Where the law where you are requires people to be told that their speech is being translated, captioned or kept, that notice is your responsibility. You are the one standing in the room.

9What happens to the audio

The audio of your event is processed as it flows and is never written to a disk. There is no recording and no archive of what your room sounded like.

The text is kept for as long as your venue’s setting says — three, six or twelve months — and then deleted automatically. Nothing from your events is used to train or evaluate any model. Our privacy policy sets all of this out in full.

Because the service transmits what is said to whoever holds the link, think about what is being spoken before you start. If part of your event is confidential, stop the session for that part.

10When it gets something wrong

Tell us. A specific example — the venue, roughly when, what was said and what appeared — is worth far more to us than a general impression.

An example you send us in a support message is content you have chosen to send, and we read it to find and fix what you reported. That is a different thing from your event content, which is never used to train or evaluate a model. We keep a support message only for as long as it takes to resolve what you told us about.

Use the contact form. If the same thing fails repeatedly in the same language, say so plainly; a pattern is more useful to us than a single line.