Why an AI guesses when it has never heard a name
A speech AI has learnt from vast amounts of text and recordings how words sound and what they mean. When it hears a name for the first time, it picks the word that sounds closest. A speaker called Anna Muster can turn into a “pattern” in English, and an abbreviation into a different word altogether.
Technical terms raise a similar problem. In your field many words have a precise meaning and an established translation. Without preparation, the AI often chooses the translation that is most common in everyday language.
Interpreters have always dealt with this by preparing. They read the programme and the documents and build a glossary. In AI interpreting the glossary takes on that role. It tells the AI before the event which names and terms will come up and what they are called in each language.
What goes into a glossary for interpreters
The glossary holds everything an unprepared listener could misunderstand or mistranslate. Everyday words do not need an entry. The examples come from a fictional conference on building and energy.
- Speakers’ namesWrite them as they appear in the programme, with title and role, for example “Dr Anna Muster, Head of Research”. Mark names that are also ordinary words as names.
- Organisations and authoritiesThey go in with their official name in each language. The Grosser Rat, a cantonal parliament, is Grand Conseil in French and Grand Council in English.
- AbbreviationsEach abbreviation comes with its full form and its equivalent in each language. VAT is MWST in German and TVA in French.
- PlacesPlaces with more than one name matter most. Fribourg is Freiburg in German, and a short note makes clear that the Swiss city is meant.
- Products and projectsProduct and project names usually stay the same in every language, such as a planning software called “Musterplan”. Without an entry the AI may translate the name.
- Technical termsThese are words with a fixed meaning in your field. Embodied energy is “graue Energie” in German and “énergie grise” in French.
- One column per target languageEach language you interpret into gets its own column with the right form. PONTAVOX imports such tables from Excel, Word or a CSV file.
Background documents give the context
Alongside the glossary come documents such as the programme, the speaker list and the text of the slides. PONTAVOX takes up to about 36,000 characters of them. If you have more, give priority to the talks with the most technical terms.
These documents are kept apart as data, so that a sentence inside them does not act as an instruction. A line such as “Please summarise briefly” on a slide should not change the interpretation.
Confidential papers do not belong in the documents. Terms and context are all the interpretation needs. More on this under data protection in AI interpreting.
What a glossary cannot do
A glossary helps a great deal, but it has limits.
- How names are pronouncedThe glossary helps the AI recognise a name. We cannot promise how the voice will pronounce it.
- A glossary from another eventIt does more harm than good. The AI then expects names and terms that will not come up that day and may match similar-sounding words wrongly. That is why every event gets its own glossary.
- The unexpectedAnything that arises in the hall is in no glossary, such as a heckle, a name called out from the audience or a pun. For debates with frequent interruptions, an interpreting team is the better choice.
- The second pathIf rebuilding a language after a fault fails twice, a second path takes over. It runs without the glossary. After three minutes PONTAVOX returns to the first path at the end of a sentence.
Measured: 9 out of 10 technical terms right
At the dress rehearsal on 23 October 2026 we counted. With the glossary, 9 out of 10 technical terms and 3 out of 3 names came out right.
The count comes from a single day, so it is no promise for every event. It does show where the work lies: one term in ten went wrong. That is why a trial in the hall with a page from your programme is worth arranging. Before the start, the Studio also asks whether the glossary and the documents are in order.
How we prepare the glossary at the Eventforum Bern
For now PONTAVOX is only available for events at the Eventforum Bern. In the booking enquiry, choose the interpretation equipment and switch on the PONTAVOX option. You will receive a quote within 12 hours, and we then prepare the glossary with you.
- 1. You send the documentsYou send us the programme and the speaker list, plus your own list of names and technical terms if you have one.
- 2. We enter themWe put names, organisations, abbreviations and technical terms into the glossary, one column per target language, and add the documents.
- 3. A trial by arrangementIn the hall we test a page from your programme before your guests arrive.
- 4. The Studio checksBefore the start, the PONTAVOX Studio asks whether the glossary and the documents are in order. What else PONTAVOX does is described under Features.
- 5. During the eventAs the event goes on, PONTAVOX is told the name of the next speaker.
- 6. After the eventIf you wish, you receive a one-page A4 report with time on air, availability, dropouts, distance to the speech and the glossary, without the transcript.
Interpreter glossary template
You can copy this template into Excel or Word. The first column holds the term as it appears in the programme or on the slides. Then comes the right form for each language, and finally a note for everyone who works with the list. For more languages, add one column per language.
How to plan such a conference as a whole is covered in the guide Planning a multilingual conference.
| Term | German | French | English | Note |
|---|---|---|---|---|
| Dr Anna Muster | Dr. Anna Muster | Dr Anna Muster | Dr Anna Muster | Speaker, block 2; a name, never translate |
| Grand Council | Grosser Rat | Grand Conseil | Grand Council | Cantonal parliament |
| VAT | Mehrwertsteuer (MWST) | taxe sur la valeur ajoutée (TVA) | value added tax (VAT) | Abbreviation differs by language |
| Fribourg | Freiburg | Fribourg | Fribourg | The Swiss city |
| embodied energy | graue Energie | énergie grise | embodied energy | Technical term |
| Musterplan | Musterplan | Musterplan | Musterplan | Software; the name stays |
Examples from a fictional conference; the people and products are invented.