Dutch is spoken as a first language by more than 20 million people, mainly in the Netherlands, Flanders and Suriname, and shares one standard written form maintained by the Dutch Language Union. For speech recognition it brings three characteristic problems: long compound words that make word error rate volatile, verb endings and vowel spellings that sound identical but are written differently, and heavy borrowing from English in business speech. Flemish adds a second pronunciation and vocabulary norm on top.
Key takeaways
- Dutch writes compounds as one word. A single wrong boundary turns one reference word into three errors.
- "Gebeurd" and "gebeurt" sound identical. Choosing correctly needs grammar, not acoustics.
- Netherlands and Flemish Dutch differ in pronunciation, vocabulary and informal register; test on both if you serve both.
- English loanwords take Dutch grammar: "gedownload", "geüpdatet", "de deadline halen".
- Numbers are said units first ("vierentwintig") and "half drie" means 2:30.
The language and its varieties#
Standard Dutch spelling is set jointly for the Netherlands, Flanders and Suriname by the Nederlandse Taalunie (Dutch Language Union), which publishes the official word list, the Woordenlijst Nederlandse Taal, known as the Groene Boekje. A transcript has one correct spelling regardless of where the speaker is from.
Pronunciation varies far more than spelling. The g is a harsh, velar or uvular sound in much of the Netherlands and a soft, palatal one in Limburg, Brabant and Flanders. Flemish speakers keep some vowel distinctions that many Netherlands speakers have merged, and the rhythm of Flemish speech differs noticeably. In Flanders, much informal speech is in tussentaal, a register between dialect and standard, with its own pronouns and verb forms.
Vocabulary differs too. A Flemish speaker says gsm for a mobile phone, goesting for appetite or desire, and ambetant for annoying, words a model trained mostly on Netherlands speech may not know. West Frisian, spoken in Friesland, is a separate language, not a variety of Dutch.
See accents and regional speech for how to measure gaps between varieties.
Compound words#
Dutch, like German, writes compounds as a single word: ziekenhuisopname (hospital admission), zorgverzekeraar (health insurer), arbeidsongeschiktheidsverzekering (disability insurance). Compounds are formed freely, so many never appear in any dictionary or training corpus, and they often contain a linking morpheme: the -s- in arbeidsovereenkomst, the -en- in pannenkoek.
For recognition, the problem is less about hearing the compound than about writing it as one word. For scoring, a split compound is catastrophic:
This is why Dutch evaluations should report character error rate alongside WER, and why the choice of text normaliser matters. More on this in compound words and morphology.
Spellings that sound cannot decide#
Several Dutch spelling distinctions have no audible counterpart.
Verb endings in d, t and dt#
Final d is pronounced /t/ in Dutch, so word, wordt and wort sound the same. The correct form depends on the subject and tense: ik word, hij wordt. Past participles cause the same trouble:
het is gisteren gebeurd vs het is gisteren gebeurt
Only the grammar of the sentence, here the auxiliary is and the word gisteren, shows that the participle gebeurd is meant. This is a classic error for Dutch writers too, and a good test of a model's language modelling.
ei and ij#
The digraphs ei and ij represent the same sound in standard pronunciation. Leiden (to lead, and the city) and lijden (to suffer) are homophones, as are reizen (to travel) and rijzen (to rise). Context usually resolves them; with names and rare words it may not.
au and ou#
Au and ou are likewise pronounced alike, as in rauw (raw) and rouw (mourning).
English loanwords and code-switching#
Business Dutch borrows heavily from English: meeting, deadline, feedback, call, update. Borrowed verbs take Dutch morphology, and the Taalunie sets the spelling: gedownload, geüpdatet, geforward. A model has to recognise an English stem with Dutch pronunciation and wrap it in Dutch affixes.
heb je de slides al gecheckt? vs heb je de slights al gecheckt?
Longer English phrases also appear inside Dutch sentences, especially in tech, finance and consulting. See code-switching and European English.
Numbers and formatting#
Dutch writes a decimal comma and a thousands point: drie komma vijf miljoen becomes 3,5 miljoen, and 12.500 means twelve thousand five hundred. Dates run day-month-year. Postcodes have four digits and two letters (1234 AB), and phone numbers are usually grouped by area code. These conventions belong in the inverse text normalisation layer and in your evaluation rules. See numbers and alphanumerics.
Data and evaluation#
Dutch is well resourced for a mid-sized language. The main corpora:
| Corpus | Content | Notes |
|---|---|---|
| Spoken Dutch Corpus (CGN) | Around nine million words, Netherlands and Flanders | Many styles, including spontaneous conversation and telephone speech |
| JASMIN-CGN | Children, older adults, non-native speakers | Extends CGN to groups that recognisers handle poorly |
| Common Voice (nl) | Volunteer read sentences | Varied microphones; read, not conversational |
| Multilingual LibriSpeech (Dutch) | Audiobooks | Read literature |
| FLEURS (nl) | Read translated sentences | Parallel with other languages; common benchmark |
The N-Best 2008 evaluation was the first large benchmark campaign for Dutch ASR, covering both Netherlands and Flemish speech in broadcast and conversational telephone conditions. More recently, Feng and colleagues (2021) used CGN to show that a Dutch recogniser performed worse for non-native speakers, children, older adults and some regional accents. For production, build a test set from your own recordings, with both varieties if you serve both; see building your own ASR test set.
How Reson8 approaches this#
Reson8 is built in Amsterdam, and Dutch is one of its nine core languages. On FLEURS, Resonant-1 scored 4.88% WER on Dutch, the lowest in Reson8's published comparison for that language; see the benchmarks page. FLEURS is read speech, so evaluate on your own Dutch and Flemish recordings as well.
Dutch domain terms, such as clinical compounds or product names, can be supplied through text-only domain adaptation with no audio or retraining. Processing runs on EU GPUs with zero audio retention. See the documentation.
Common questions#
Is Flemish a separate language for ASR?
No. It is written with the same standard spelling, so one model can serve both. Pronunciation and vocabulary differ enough that you should test on Flemish audio separately if you serve Belgian users.
Why is Dutch WER often higher than Spanish WER for the same system?
Partly because compounds and d/t endings create word-level errors that Spanish, with more transparent spelling and fewer compounds, does not. Compare CER as well as WER.
Should English loanwords be spelled the English way?
Follow the Taalunie word list for established loans, such as geüpdatet. Decide your rule for newer terms in advance and apply it to references too.
Does Dutch ASR handle Frisian?
West Frisian is a distinct language. A Dutch model will usually transcribe it poorly or render it as Dutch.
References#
- Oostdijk, N. (2000). The Spoken Dutch Corpus: Overview and First Evaluation. LREC.
- Cucchiarini, C., Van hamme, H., van Herwijnen, O., & Smits, F. (2006). JASMIN-CGN: Extension of the Spoken Dutch Corpus with Speech of Elderly People, Children and Non-natives in the Human-Machine Interaction Modality. LREC.
- van Leeuwen, D. A. et al. (2009). Results of the N-Best 2008 Dutch Speech Recognition Evaluation. Interspeech.
- Feng, S., Kudina, O., Halpern, B. M., & Scharenborg, O. (2021). Quantifying Bias in Automatic Speech Recognition. arXiv:2103.15122.
- Nederlandse Taalunie. Woordenlijst Nederlandse Taal. woordenlijst.org.
- Conneau, A. et al. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. IEEE SLT.