Europe has 24 official EU languages and dozens more regional and minority languages, spread across Germanic, Romance, Slavic, Finno-Ugric, Baltic, Celtic, Greek and Semitic families. Speech recognition quality across them is uneven, because training data is uneven: English dominates, a handful of large languages follow, and the rest have far less. For business use, the question is not how many languages a system lists but how deeply it handles the ones you need, from number formats to regional accents to your domain's terms.
Key takeaways
- A long language list says little about quality. Accuracy in the tail of a hundred-language model can be far below its headline languages.
- Depth means correct formatting, handling of varieties such as Flemish or Swiss French, domain vocabulary, and published evaluation per language.
- Public European speech data comes mostly from read speech, audiobooks and parliament, which differs from calls and meetings.
- Each language has its own hard problems: compounds in Dutch and German, homophones in French, morphology in Polish.
- For European organisations, where audio is processed matters as much as accuracy. See EU data residency.
Europe's linguistic diversity#
The EU's 24 official languages range from German, with the most native speakers in the Union, to Maltese and Irish, with far fewer. Beyond them sit regional languages such as Catalan, Basque, Welsh, Frisian and Luxembourgish, and the languages of large immigrant communities, including Turkish and Arabic. Many European workplaces also run on English spoken as a second language, covered in European English.
The languages differ in ways that matter for recognition:
- Word formation. Dutch, German and Swedish write compounds as one word. Polish and other Slavic languages inflect nouns for case. Both multiply the number of distinct word forms a model must handle. See compound words and morphology.
- Spelling and sound. Spanish and Italian spelling is close to pronunciation. French and English spelling is not, and French has many homophones that only grammar can separate.
- Prosody. Swedish uses pitch accent to distinguish words. Stress is fixed in Polish and free in Russian and Italian.
- Locale conventions. Decimal commas, 24-hour times, day-month dates and currency placement differ from English defaults. See numbers and alphanumerics.
The data imbalance#
Most large speech models are trained on data dominated by English. The large European languages, German, French and Spanish, have substantial resources. Mid-sized languages such as Dutch, Polish, Portuguese and Swedish have less, and smaller languages much less. The META-NET white papers (2012) assessed language technology support across European languages and found that most had weak or fragmentary support; the European Language Equality project revisited the question a decade later and found large gaps remained.
The main public speech corpora for European languages each have a particular character:
| Corpus | Source | Coverage | Character |
|---|---|---|---|
| Common Voice | Volunteers reading prompts | Most European languages | Read sentences, varied microphones |
| Multilingual LibriSpeech (MLS) | LibriVox audiobooks | English, German, Dutch, French, Spanish, Italian, Portuguese, Polish | Long-form read literature |
| VoxPopuli | European Parliament | EU languages | Formal speeches and interpretation |
| FLEURS | Read translated sentences | 102 languages | Short, parallel across languages |
| Spoken Dutch Corpus (CGN) | Many genres | Dutch and Flemish | Includes spontaneous and telephone speech |
None of these looks much like a contact-centre call or a hospital ward round. That gap is why vendor results on public sets can overstate accuracy on real audio, and why building your own test set matters more for European languages than for English.
What depth means#
A system "supports" a language when it outputs words in it. Depth is everything above that.
- Formatting. drie komma vijf procent should become 3,5% in Dutch, not 3.5%. See inverse text normalisation.
- Varieties. Belgian French septante, Flemish vocabulary, European versus Brazilian Portuguese. See accents and regional speech.
- Domain terms. Clinical Dutch mixes Latin, English and Dutch. German legal language has long, precise compounds.
- Evaluation. A language-specific text normaliser and published results per language, not only an average across all of them.
Language-specific challenges#
| Language | Main challenges |
|---|---|
| Dutch | Compounds, Flemish variety, English loanwords, homophonous verb endings |
| German | Long compounds, case endings, Swiss and Austrian varieties |
| French | Liaison, silent inflections, homophones, regional numbers |
| Spanish | Peninsular and Latin American variants, fast speech |
| Italian | Regional accents, dialect influence, speaking rate |
| Polish | Rich inflection, consonant clusters |
| Portuguese | European and Brazilian variants, vowel reduction |
| Swedish | Pitch accent, compounds, smaller data pool |
Speakers also switch between languages, particularly between a local language and English in business settings. See code-switching and multilingual speech recognition.
How Reson8 approaches this#
Reson8 focuses on nine European languages in depth rather than a long list: Dutch, English, French, German, Italian, Polish, Portuguese, Spanish and Swedish. On FLEURS, Resonant-1 averaged 4.22% WER excluding Swedish (English 4.69, German 3.83, French 4.71, Dutch 4.88, Spanish 2.67, Polish 4.56), leading in French, Dutch, Spanish and Polish. Swedish scored 7.38%, which is why it is reported separately. Full tables are on the benchmarks page.
All processing runs on EU GPUs with zero audio retention and no training on customer data. The trade-off is explicit: if you need Finnish, Greek or Hungarian, Reson8 does not cover them.
Common questions#
Is a model trained on 100 languages better than one trained on 9?
Not necessarily for the languages you need. Model capacity and data are shared across languages, and quality usually falls with a language's share of the training data. Compare per-language results.
Which European languages are hardest for ASR?
Those with little transcribed speech data are hardest in practice. Among large languages, heavily compounding and inflecting languages such as German, Dutch and Polish produce more word-level errors, and French homophones are a persistent source of spelling errors.
Do FLEURS results predict accuracy on phone calls?
Only loosely. FLEURS is read, wideband speech. Calls are spontaneous and often 8 kHz. See telephony transcription.
Why does EU processing matter for speech?
Voice recordings are personal data under GDPR, and some sectors require processing within the EU. See GDPR and voice data.
References#
- Conneau, A. et al. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. IEEE SLT. arXiv:2205.12446.
- Pratap, V. et al. (2020). MLS: A Large-Scale Multilingual Dataset for Speech Research. Interspeech.
- Wang, C. et al. (2021). VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. ACL.
- Ardila, R. et al. (2020). Common Voice: A Massively-Multilingual Speech Corpus. LREC.
- Oostdijk, N. (2000). The Spoken Dutch Corpus: Overview and First Evaluation. LREC.
- Rehm, G., & Uszkoreit, H. (Eds.) (2012). META-NET White Paper Series: Europe's Languages in the Digital Age. Springer.
- Rehm, G., & Way, A. (Eds.) (2023). European Language Equality: A Strategic Agenda for Digital Language Equality. Springer.