A tender is a strange instrument for buying this. It has to be written by somebody who has not seen the system working, evaluated by somebody who will not operate it, and it has to be specific enough that a bad supplier cannot satisfy it on paper.
Most language-access tenders fail that last test. They specify a number of languages, an accuracy percentage and a price, all three of which a weak system can satisfy while being useless in the chamber.
Here is what to specify instead.
Specify the output, not the feature
“Provides live subtitles” is satisfied by a full-screen web page on a monitor in the corner. If the sitting is produced with a video mixer, that is not the same product as captions that composite into the broadcast signal.
Ask for the mechanism by name: an output with a real alpha channel that the production mixer treats as an input, or full-screen with a chroma key, and say which your room needs. A supplier who only has a web page to key will tell you, because the alternative is failing a demonstrable requirement.
Specify the failure mode
This is the clause almost nobody writes and the one that matters most in a public setting, because a public sitting is streamed and recorded.
Require that on loss of connectivity the audience-facing output clears. Not freezes, not shows an error. A caption frozen on the big screen during a council sitting is worse than no caption at all: it is a false record, visible to the room and to the stream, and nobody watching can tell it has stopped being true.
Specify the data questions as three separate answers
“GDPR compliant” is not an answer, it is an assertion. Ask for three facts, each of which can be checked:
- Where is audio processed? A country or a region, not “in the cloud”.
- What is retained, and for how long? Ask separately about the live path and about any summarisation or minutes feature, because they are usually different answers.
- Who are the sub-processors? The full list, published, with what each one handles and where. This is already a GDPR obligation, so a supplier who cannot produce it on request has a problem that predates your tender.
Be aware that “processed in the EU” is rarely absolute across a whole pipeline. Speech recognition, translation and any summarisation may sit with different providers in different places. A supplier being precise about which is which is a good sign, not a bad one; a supplier answering with one confident word is worth a follow-up question.
Specify accessibility evidence, not an accessibility claim
A percentage accuracy figure quoted in a bid is unverifiable and depends entirely on the audio you feed it. What can be verified after the fact is usage: how many devices followed each session and in which languages. Require a per-session report, and you have turned an unfalsifiable claim into something your own records can settle.
Two things to be careful about in the other direction. If you require a formal conformance statement against a specific standard, ask bidders directly whether they issue one. Some will say yes because a form asked, and a supplier who tells you plainly that they do not yet issue one is giving you better information than a tick in a box. And separately, a transcript is not accessibility: a published video with no captions on it is still a published video with no captions on it, whatever the supplier delivered alongside.
Four questions for the clarification round
- Does the output composite into our existing production, and by what mechanism?
- What appears on screen during a thirty-second loss of connectivity?
- Where is audio processed, what is retained, for how long, and who are the sub-processors?
- What per-session evidence of use do we receive, and is it included in the price?
If those four have clear answers, the rest of the specification is mostly price. If any of them does not, no amount of specifying language counts will save the delivery.