Most English spoken in European workplaces is spoken by people for whom it is a second or third language. A meeting in Amsterdam, a support call from Warsaw or a board call with French and German directors all run in English with accents shaped by the speakers' first languages. Speech recognition models are mostly trained and benchmarked on native British and American English, so European English is a systematic blind spot: error rates are higher, and the headline benchmark numbers do not show it.
Key takeaways
- Non-native accents follow predictable patterns from the speaker's first language: final devoicing for Dutch and German speakers, dropped h for French speakers, an added vowel before s-clusters for Spanish speakers.
- Standard English benchmarks (LibriSpeech, TED-LIUM, GigaSpeech) are dominated by native speakers. Public sets with labelled non-native accents are small.
- European English mixes in local names, places, institutions and false friends, which are rare-word problems as much as accent problems.
- Measure the accent gap directly: slice your test set by speaker background and compare, rather than relying on a single average.
Why European English is its own problem#
English is the working language of many European companies, EU institutions, research and most cross-border customer service. The English in those settings has three properties that general English ASR handles less well than native speech:
- Accent: vowels, consonants, stress and rhythm carried over from the speaker's first language.
- Vocabulary: local names, streets, companies, laws and institutions, often pronounced the local way in the middle of an English sentence.
- Usage: calques and false friends that are normal in European English but rare in the training data, such as actual meaning "current", eventually meaning "possibly", or planning as a noun meaning "schedule".
The first is an acoustic problem. The second and third are language-model problems: the model hears the sounds but assigns low probability to the word sequence, and substitutes something more familiar.
Accent patterns by first language#
Second-language accents are not random. Each first language brings a characteristic set of substitutions, which is why a model that has seen some German-accented English generalises to other German speakers.
| First language | Common feature | Example said | Example heard |
|---|---|---|---|
| Dutch, German | Final devoicing | bed, dog | bet, dock |
| German | w and v merged | wine | vine |
| Dutch, German, French, Italian | th as /s/, /z/, /t/ or /d/ | think, three | sink, tree |
| French | Dropped or added h; final-syllable stress | hair, heat | air, eat |
| Spanish, Portuguese | Vowel added before s-clusters | Spain, student | espain, estudent |
| Spanish, Italian, Polish | Short and long vowels merged | ship, full | sheep, fool |
| Polish, other Slavic | No article system; w as /v/ | the report | report |
| Swedish | j and y sounds exchanged; /z/ devoiced | jam, zoo | yam, soo |
Many of these produce real English words, which is why they are dangerous. A model hearing we need to sink about this will usually correct it, because sink about is improbable. The bed is booked heard as the bet is booked is not corrected, because both are plausible. Minimal pairs in business speech, such as fifteen and fifty, sixteen and sixty, become far riskier when stress placement shifts: speakers of languages with fixed or different stress often stress FIFteen like FIFty.
Prosody and rhythm#
English is stress-timed, with reduced unstressed syllables (to as /tə/). Speakers of syllable-timed languages such as French, Spanish and Italian often give every syllable similar weight, and pronounce function words in full. The model hears clear, unreduced to, for and of, which sounds like emphasis in native speech and can nudge it towards two, four and off. Intonation carried over from the first language also affects endpointing and end-of-turn detection: a falling pitch that ends a turn in one language can mark mid-sentence continuation in another.
Local names and mixed vocabulary#
A German customer on an English support call will still say their street as Königsberger Straße, their insurer as Techniker Krankenkasse and their city as München, not Munich. A Dutch colleague will refer to the Belastingdienst or a BSN. These insertions are a form of code-switching, and an English-only model will force them into English-sounding words: Belastingdienst heard as blasting deans.
This is a rare word and proper noun problem more than an accent problem, and it is often the larger source of meaningful errors. Context biasing with local names, addresses and institutions can do more for European English accuracy than accent-specific acoustic training. See custom vocabulary and context biasing.
Measuring the accent gap#
Standard English benchmarks tell you little here. LibriSpeech is audiobook reading by mostly American volunteers; TED-LIUM, GigaSpeech and SPGISpeech are mostly native speakers in talks, podcasts and earnings calls. Sets that do include non-native English:
- AMI Meeting Corpus: meetings recorded in Edinburgh, Idiap and TNO, with a substantial share of non-native participants. It is part of the Open ASR Leaderboard and its error rates are consistently among the highest on that board.
- VoxPopuli: European Parliament speeches, including English from MEPs of many nationalities.
- Edinburgh International Accents of English Corpus (EdAcc): conversational English from speakers with a wide range of first languages and accents, released specifically to measure this gap.
- L2-ARCTIC: read speech from non-native speakers, but its first languages are Arabic, Mandarin, Hindi, Korean, Spanish and Vietnamese, so its coverage of European accents is limited to Spanish.
- Common Voice: English clips carry optional self-reported accent labels, useful for slicing.
- Speech Accent Archive (George Mason University): the same paragraph read by speakers of hundreds of first languages. Too small for training, useful for listening tests.
Accent gaps in ASR are documented beyond Europe. Koenecke et al. (2020) found substantially higher error rates for Black American speakers than white American speakers across five commercial systems, showing that aggregate WER can conceal large group differences.
How Reson8 approaches this#
Reson8's published English results are on public sets: a 6.51% average across the Open ASR Leaderboard short-form sets, including AMI at 9.39%, and 4.69% on FLEURS English. These sets are mostly native speech, and Reson8 has not published a separate figure for non-native European English. Resonant-1 supports in depth nine European languages, which are the first languages of many of the people speaking English in European workplaces. For local names and institutions, text-only domain adaptation lets you supply those terms through the API.
Common questions#
Is non-native English always harder for ASR?
Usually, because training and benchmark data are dominated by native speakers. The size of the gap depends on the accent, the speaker's fluency and how much similar speech the model has seen.
Should I use a multilingual model for European English?
Often yes. It handles embedded local names better and has seen the phonology of the speakers' first languages. Test both on your own audio.
Do accent-specific models help?
They can, but they fragment your stack and require knowing the accent in advance. For most deployments, broader training data plus context biasing for local vocabulary is more practical.
Which public benchmark best reflects European workplace English?
AMI and VoxPopuli are the closest widely used sets, and EdAcc for conversational accented speech. None substitutes for recordings of your own users.
References#
- Jenkins, J. (2000). The Phonology of English as an International Language. Oxford University Press.
- Koenecke, A. et al. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14).
- Sanabria, R. et al. (2023). The Edinburgh International Accents of English Corpus: Towards the Democratization of English ASR. ICASSP 2023.
- Carletta, J. et al. (2005). The AMI Meeting Corpus: A Pre-Announcement. Machine Learning for Multimodal Interaction (MLMI).
- Zhao, G. et al. (2018). L2-ARCTIC: A Non-Native English Speech Corpus. Interspeech 2018.
- Weinberger, S. Speech Accent Archive. George Mason University.
- Wang, C. et al. (2021). VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. ACL 2021.