Swedish combines three things that make speech recognition harder: a pitch accent that distinguishes otherwise identical words, compounds written as one word whose meaning changes if they are split, and a smaller pool of training data than the large European languages. It is also a language with strong regional variation, including Finland Swedish, and close neighbours in Norwegian and Danish that speakers mix in. Swedish is Reson8's weakest published FLEURS result.
Key takeaways
- Swedish has two word accents. Anden with accent 1 is "the duck"; with accent 2 it is "the spirit". Spelling is identical, so pitch matters mainly for segmentation and compound detection.
- Compounds are written as one word, and splitting them changes the meaning: sjuksköterska is a nurse, sjuk sköterska a sick nurse.
- Around ten million speakers means less open data than for German, French or Spanish, and less conversational data in particular.
- Formatting uses a space for thousands, a decimal comma and kr after the amount; ISO dates (2026-09-23) are standard.
- Reson8's published FLEURS Swedish WER is 7.38%, higher than its other published languages and excluded from its 4.22% multilingual average.
What makes Swedish hard#
Pitch accent#
Most Swedish dialects distinguish two tonal word accents. Accent 1 has a single pitch movement; accent 2, typical of compounds and many two-syllable words, has a characteristic two-peaked contour in Central Swedish. Minimal pairs include anden (the duck / the spirit), tomten (the plot of land / the Christmas gnome) and buren (the cage / carried).
Because both members of each pair are spelled the same, pitch accent rarely causes a written error on its own. Its importance for ASR is indirect. Accent 2 is the prosodic signal of a compound, so it tells a listener that sjuksköterska is one word and not sjuk sköterska. Pitch contours also differ strongly by region: the accents are realised differently in Stockholm, Gothenburg, Skåne and Dalarna, and Finland Swedish largely lacks the distinction. A model that relies on one region's melody for word segmentation will stumble on another's.
Compounds and särskrivning#
Like German and Dutch, Swedish writes compounds as one word, and new ones are formed freely: kundtjänstmedarbetare (customer service agent), försäkringskassan, sjukskrivningsintyg. Writing them as separate words, särskrivning, is a recognised spelling error in Swedish because it often changes meaning:
| Compound | Meaning | Split | Meaning |
|---|---|---|---|
| sjuksköterska | nurse | sjuk sköterska | sick nurse |
| rökfritt | no smoking | rök fritt | smoke freely |
| kassapersonal | checkout staff | kassa personal | lousy staff |
| brunhårig sjuksköterska | brown-haired nurse | brun hårig sjuk sköterska | brown, hairy, sick nurse |
For ASR, a split compound costs two errors under WER, and it may also be a genuine meaning error. The linking -s- (försäkringsbolag) is short and often lost. See compound words and morphology.
Vowels, retroflexes and the sj-sound#
Swedish has nine vowel qualities, each with a long and short form, and quantity is contrastive: vit (white, common gender) versus vitt (white, neuter). The difference is a single t in writing and mostly vowel length in speech, and it marks grammatical agreement.
In Central Swedish, r merges with a following dental consonant into a retroflex: barn, kort, fors. This also happens across word boundaries, so har du sounds like one unit with a retroflex d. Southern Swedish uses a uvular r and has no retroflexes at all. The sj-sound in sju (seven), sjuk, skjorta and stjärna is notoriously variable across speakers and regions, and in some varieties approaches the tj-sound of kjol, tjugo. Spelling for this sound is also varied (sj, skj, stj, sk, sch), which is a source of errors on rarer words.
Dialects and neighbouring languages#
- Central Swedish (Stockholm, Mälardalen): the reference for most broadcast and training data.
- Southern Swedish (Skåne): uvular r, diphthongised long vowels, a distinctive melody, and no retroflexes.
- Gothenburg and western Swedish: different accent 2 contours and local vocabulary.
- Norrland and Gotland: distinct prosody and vowels.
- Finland Swedish: an official language of Finland, with no tonal word accents, different vowel quality and Finnish-influenced rhythm. It is standard Swedish in writing but sounds noticeably different.
Swedish, Norwegian and Danish are partly mutually intelligible, and in Scandinavian workplaces speakers commonly use their own language with each other. A Norwegian speaking in a Swedish meeting will be forced into Swedish spellings by a Swedish-only model. See code-switching and language identification.
Numbers, dates and formatting#
Swedish conventions for inverse text normalisation:
| Spoken | Written | Note |
|---|---|---|
| tolvhundrafemtio kronor och femtio öre | 1 250,50 kr | Space for thousands, decimal comma, kr after |
| tjugotredje september tjugohundratjugosex | 2026-09-23 or 23 september 2026 | ISO date order is standard in Sweden |
| halv tre | 14.30 or 2.30 | halv tre is half past two; both 14.30 and 14:30 are seen |
Number words are written as one word (tjugotre, tvåhundrafemtio), and years are often read in hundreds: nittonhundranittiotvå for 1992, tjugohundratjugosex or tvåtusentjugosex for 2026. Swedish personal identity numbers (personnummer, YYMMDD-XXXX) are dictated constantly in healthcare and public services, and are a good candidate for format-aware transcription. See structured identifiers.
Data availability and public test sets#
With around ten million speakers, Swedish has much less open speech data than the large European languages, and the gap is widest for conversational and telephone speech. Available resources include:
- FLEURS (Swedish split): read Wikipedia sentences, parallel with other languages.
- Common Voice (
sv-SE): smaller than the German or French subsets. - NST Swedish: a read-speech corpus originally produced by Nordisk Språkteknologi, now distributed by the National Library of Norway's Språkbanken.
- VoxPopuli: European Parliament speech.
- The National Library of Sweden (KBLab) has published Swedish speech models trained on large collections of Swedish broadcast audio and parliamentary recordings, which shows how much there is outside public benchmarks.
How Reson8 approaches this#
Swedish is one of the nine languages Resonant-1 supports. Its FLEURS Swedish WER is 7.38%, clearly higher than its other published languages (2.67% to 4.88%), and Reson8 reports its 4.22% multilingual FLEURS average excluding Swedish, with the Swedish figure listed separately in the benchmarks post. Swedish is the weakest of its published results, and anyone evaluating Resonant-1 for Swedish should test on their own audio.
Common questions#
Does pitch accent cause transcription errors?
Rarely directly, because accent pairs share a spelling. It matters for compound detection and segmentation, and regional differences in pitch confuse models trained on one variety.
Why is Swedish WER higher than German or Spanish WER?
Mainly less training data, especially conversational data, plus compounding that makes WER count boundary errors twice. Regional variation adds to the gap.
Can a Swedish model transcribe Norwegian or Danish?
It will produce Swedish-looking text for Norwegian and much worse output for Danish. Use language identification or a model that supports each language.
References#
- Bruce, G. (1977). Swedish Word Accents in Sentence Perspective. Lund: Gleerup.
- Riad, T. (2014). The Phonology of Swedish. Oxford University Press.
- Conneau, A. et al. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. arXiv:2205.12446.
- Ardila, R. et al. (2020). Common Voice: A Massively-Multilingual Speech Corpus. LREC 2020.
- Wang, C. et al. (2021). VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. ACL 2021.
- Malmsten, M., Haffenden, C., & Börjeson, L. (2022). Hearing voices at the National Library: a speech corpus and acoustic model for the Swedish language. arXiv:2205.03026.