Does an AI receptionist understand Dutch? Yes, bar one thing

Yes. An AI receptionist understands everyday Dutch well, and postcodes and addresses mostly too. Names and numbers are more error-prone, and that is where a check belongs.

That is the whole answer. Below is why it works that way, what the figures doing the rounds do and do not tell you, and how to test it yourself in five minutes.

Officially Dutch is easy, yet it has the most rules that do nothing

The US State Department ranks Dutch among the easiest languages to learn. Yet it has more rules that add no meaning than 21 other languages.

The Dutch like to say their language is one of the hardest in the world. The Foreign Service Institute, which tracks how long each language takes American diplomats, disagrees. Dutch sits in category one, 24 weeks, alongside Danish and Italian. Arabic, Chinese, Japanese and Korean sit in category four, 88 weeks.

Linguist Sterre Leufkens of Utrecht University did find something else. She compared 22 languages on how often a rule has a form but no meaning. Dutch was the outlier: two words for “the” with no logic behind which is which, two object pronouns that even natives mix up, three letters e in one word that all sound different. She calls those rules male nipples. They are there, and they do nothing.

For an AI receptionist neither point matters. It never has to pick an article or conjugate a verb. It has to hear what you said. And that is a different kind of hard.

Everyday sentences are not the problem

A sentence like “I want to move my Thursday appointment” is understood correctly almost every time. The technology has been ready for that for years.

Dutch sits among the best supported languages for speech recognition, together with English, German and French. Vendors quote accuracies of 85 to 95 percent on clear audio, and in benchmarks the word error rates sit below ten percent.

Two caveats on those figures. They are about whole sentences, and a sentence lets the system fill in from context: hear “appointment” and “Thursday” and “move” almost suggests itself. And most benchmarks use read-aloud text in studio quality, not someone saying their surname on a poor line. The figure is right, but it does not measure what you want to know.

Names and numbers are more error-prone

With a surname, a lettered house number or a phone number there is no context to guess from. That is where the few errors sit.

Not many, but exactly the ones that matter. A few things that are harder on a Dutch phone call than an ordinary sentence:

• Surname prefixes. Is it “Van der Meulen” in three words, “Vandermeulen” in one, or “Van Dermeulen”? All three exist.
• Postcodes. Dutch postcodes end in two letters. The digits go fine. Among the letters, B, P, D and T sound alike, and so do M and N. A phone line cuts off exactly the high frequencies where the difference lives.
• House number suffixes. 8A, 45-2, 112 hs. A person hears that as one unit, a system has to learn it.
• Numbers in groups. “Zero six, twelve, thirty-four” is different from “zero six one two three four”.
• Self-correction. “Zero six twelve, no sorry, zero six thirteen.” Which one counts?
• An English word in between. “I’m calling about my appointment tomorrow”, said in Dutch with one English word. Switching language mid-sentence.

The good news: most of this now simply works, postcodes included. The point is not that it fails. The point is that these are the places where it still can, and so you want to know what the system does when it is unsure.

A name fixed as text can no longer be checked

Most AI receptionists turn speech into text first and only then into an answer. Whatever is wrong after that first step never recovers.

Speech becomes text, a language model works out a reply, and that becomes a voice again. That is why names are the weak spot, and not the language. Once “Ruijgrok” sits in the text as “Reigrok”, everything after that works with Reigrok. Nobody listens to the recording again. There is another way to build it, with one model working directly from speech to speech without the detour through text. How that difference works is covered in the piece on what an AI receptionist is. For this piece one thing is enough: ask a vendor how it is built, and what happens to a name it is not sure about.

Reading back or not: both have a price

One AI receptionist reads a name back and asks whether it is right. Another deliberately does not. Both are defensible, both have a downside.

The second school reasons that a mispronounced name sounds unprofessional. Reading back costs a few seconds and can sound wooden, especially with a name the voice does not know. But the error surfaces while the caller is still on the line and can correct it on the spot. Not reading back sounds smoother, but the error only surfaces when someone calls back and uses the wrong name, or dials the wrong number.

Whichever it is, you want one thing: that the doubt is visible to whoever reads the message. A name the system was not sure about belongs in the summary flagged as such. Then the colleague who calls back knows to check.

Flemish and dialect: ask, and test with your own callers

Standard Dutch from the Randstad is recognised best, because most training data comes from there. The stronger the accent, the more it varies per vendor.

Flemish, Brabantian and Limburgish are recognised better and better, but how well exactly varies per system and is rarely published. An honest vendor says so. If many of your callers are from Flanders or from a region with a strong accent, the only way to know is to have someone from there call before you sign.

Call it yourself with these five

You do not need a benchmark. Call a vendor’s test number with five things that are error-prone, then read the summary.

1. A surname with a prefix: “this is Van der Meulen”.
2. A postcode with similar-sounding letters: “5104 JB”.
3. A house number with a letter: “Kerkstraat 8A”.
4. A mobile number in pairs, at normal speed.
5. Correcting yourself halfway: “zero six twelve, no sorry, zero six thirteen”.

Do not only note what went right. Note what the system did when it was unsure: did it ask you to repeat, did it read back, or did it just carry on. More things to call about are in how to test an AI receptionist. And bear in mind that this is one of the five cons that disappear with a proper setup, not a reason to drop the idea.

How BeepSweep does it

BeepSweep uses one model that works directly from speech to speech, without a text detour. That is a design choice, not a promised percentage.

The connection is wideband, which lets through more of the sound than ordinary telephony. It reads names, phone numbers and addresses back until the caller confirms, and you can set it to send the caller a WhatsApp confirmation afterwards. What we see in practice: Dutch comprehension is good, postcodes included, and the small errors that remain sit exactly where this piece predicts them. That is an observation, not a measurement. It works in six languages, with Dutch first.

To hear it with your own name and address, book a demo. Questions are welcome on WhatsApp at +31 85 083 5355. For the other providers, see the comparison of ten Dutch AI phone assistants.

Frequently asked questions

Does an AI receptionist understand Flemish?

Usually, but how well varies per system and is rarely published. Have someone from Flanders call the test number before you decide.

What if it misheard my name?

A good system reads the name back and lets you correct it. If it does not, the doubt should at least be flagged in the summary, so whoever calls back checks it.

Does it work with an accent?

With a light accent it works fine. With a strong accent it depends on what the system was trained on. Testing with your own callers is the only way to know.

Can it handle Dutch and English mixed?

Most systems recognise a language switch and follow it, even mid-sentence. A single English word in a Dutch sentence is rarely a problem.