Speech is picked up live from the sound system and converted to text by speech recognition, which can then be translated automatically into other languages. The text appears within a few seconds on a screen in the room, in the livestream, or on the visitor's own phone. Some systems convert the translation back into spoken audio, so participants can listen in their own language through an earphone. Quality depends heavily on the audio supplied: a clean feed straight from the mixing desk gives far better results than a microphone picking up the room.

With clearly audible speakers, a clean audio feed and a widely used language, recognition accuracy is now high; specialist jargon, strong accents, overlapping speakers and abbreviations remain the weak spots. Many systems can be loaded in advance with a glossary of names, brands and technical terms, which noticeably improves accuracy. For plenary sessions, presentations and webinars, AI captioning is generally more than adequate and scales easily to many languages. For legally binding hearings, negotiations, and sensitive or ceremonial moments, a human interpreter remains the safe choice, because they weigh nuance, tone and context.

There are three common formats: a caption bar or separate screen in the room, a text layer in the livestream, and a web page attendees open on their own phone via a QR code. The last option is popular because everyone selects their own language and no equipment has to be handed out — though it does require reliable guest wi-fi or good mobile coverage. Where the translation is also offered as audio, participants listen through their own earbuds or through receivers handed out on site. In-room setups often combine formats, so attendees can follow along on the main screen or read privately.

The key requirement is a separate, clean audio feed from the mixing desk, either per speaker or as a combined mix. A stable internet connection is needed when processing takes place in the cloud; for sensitive meetings, a local on-site server can be used instead, so no audio leaves the building. Expect a delay of roughly two to five seconds between the spoken word and the text, noticeable, but easy to follow in practice. A test session beforehand matters too: speakers using a headset rather than a lectern microphone immediately produce better recognition.

With cloud services, the meeting audio is sent to an external server, which calls for clear agreements on data processing, storage location and retention periods under the GDPR. For closed or confidential sessions, there are solutions that run entirely within the network or on a local server, so recordings and transcripts never leave the building. It is advisable to inform participants in advance that live transcription is taking place, and to record whether the transcript will be retained or deleted afterwards. Any later use of transcripts — for minutes, subtitles or summaries — belongs in those same agreements.