Danish, Swedish, Norwegian and Finnish all have free speech data you can download today. Most of it is read sentences or parliamentary debate. Everyday conversation, the speech a voice agent, call analytics tool or meeting product actually hears, is the scarce part, and it is where these four languages are hardest for a speech model.
The short answer: use the open sets to bootstrap and as a first benchmark, then build a conversational test set for each language, with written rules for compounds, written standards and dialect forms. Train on conversation in the variety your users speak, whether you license it or record it, and report each language and region separately, with character error rate beside word error rate.
This guide covers what is available, what makes the four languages hard, how to evaluate a model on them and when licensing or collecting data makes sense. Facts about public datasets and models come from their own pages and papers, checked on 4 November 2026.
What open speech data exists for the Nordic languages
These are the main public sets, grouped by corpus. Sizes and terms are as each publisher states them.
| Corpus | Languages | Speech | Size | Terms |
|---|---|---|---|---|
| Common Voice Scripted Speech 27.0 | All four | Read sentences | Danish 16.2 h, Swedish 56.2 h, Norwegian Bokmål 2.6 h, Finnish 23.2 h | CC0 1.0, with no re-sharing and no attempts to identify speakers |
| FLEURS | All four (Norwegian Bokmål) | Read sentences from a translation benchmark | Around 10 h of training audio per language | CC BY 4.0 |
| NST ASR databases | Danish, Swedish, Norwegian | Read from manuscripts around 2000, close and distant microphone | Norwegian: about 540 h from close to 1,000 speakers | CC0 1.0 |
| CoRal v3 | Danish | Read aloud and conversation, all major dialects | 521 h read and 156 h conversation in the training split | OpenRAIL-D: commercial use allowed, but no speech synthesis or biometric identification |
| FT Speech | Danish | Folketing debates | Over 1,800 h from 434 speakers | The Folketing's open-data and Folketing TV licences |
| NPSC 2.0 | Norwegian | Storting debates from 2017 and 2018 | 140 h from 267 speakers | CC0 1.0 |
| RixVox-v2 | Swedish | Riksdag debates, protocol text aligned to the audio | Nearly 23,000 h | ODC-By 1.0 |
| Finnish Parliament ASR Corpus | Finnish | Eduskunta plenary sessions, 2008 to 2020 | Over 3,000 h from 449 speakers | Public download from the Language Bank of Finland |
| Donate Speech (Lahjoita puhetta) | Finnish | Speech donated in answer to prompts, 2020 to 2021 | About 3,200 h, about 1,600 transcribed | Paid licence for companies |
Read speech covers all four languages. Common Voice 27.0, released in September 2026, has 16.2 hours of Danish, 56.2 hours of Swedish, 23.2 hours of Finnish and only 2.6 hours of Norwegian Bokmål, with Nynorsk as a separate dataset. FLEURS records 2,009 sentences from the FLoRes translation benchmark in 102 languages, with different speakers in training and test.
The NST databases for Danish, Swedish and Norwegian were made for dictation and speech recognition by Nordisk språkteknologi, a company whose speech resources the National Library of Norway received from its bankruptcy estate in 2011. They are manuscript-read, recorded on a close and a distant microphone at once, and now in the public domain.
Parliament speech is unscripted but formal. NPSC has 140 hours from Norway's Storting, transcribed in Bokmål or Nynorsk and checked by linguists. RixVox-v2 pairs nearly 23,000 hours of Riksdag debates with text from the parliament's written protocols, aligned to the audio afterwards rather than transcribed from it. FT Speech covers the Danish Folketing, and the Finnish Parliament ASR Corpus the Eduskunta. The team behind Finland's Donate Speech corpus describes the Finnish parliament data as "rather formal in style and often read from the speaker's notes".
Everyday conversation is scarce. For Danish, CoRal is the main open source, with 156 hours of conversation in its training split and speakers of every major dialect and several non-native accents. Its licence forbids speech synthesis and biometric identification, and the download opens only after you accept its conditions. For Finnish, Donate Speech has about 1,600 transcribed hours of everyday speech, but companies need a separate paid licence from the University of Helsinki. In the table, unscripted Swedish and Norwegian come only from parliament.
Read speech doesn't teach a model conversation
Published Nordic results show why the kind of speech matters more than the number of hours.
The CoRal project's Whisper model card reports character error rate (CER) on CoRal's own test sets. The general-purpose whisper-large-v3 scores 10.1% on read-aloud Danish and 27.5% on conversation. The project's first Whisper model, fine-tuned on read-aloud data only, brought read speech down to 4.0% but scored 35.6% on conversation, worse than the model it started from. Its v3 model, trained on read-aloud and conversational data, scores 11.6% on conversation and 4.5% on read speech. The two versions differ in more than their training data, and the dataset card marks the conversational test split as tentative, but the direction is hard to miss.
In Norway, the team behind NPSC trained a model on 300 hours of NST read speech and tested it on spontaneous, dialectal speech from a separate dataset, NB Tale. Its word error rate there was 48.4%. Fine-tuning on NPSC's parliamentary speech brought it to 37.3%, a 22.9% relative improvement, and the gain was generally largest for the dialects that had done worst. The baseline ranged from 39.0% for Oslo speakers to 55.8% for Trøndelag, according to the NPSC paper. These are 2022 results from a DeepSpeech 2 model, far behind current systems, but they show the gap a read-speech corpus leaves.
Swedish shows what volume does for read speech. The National Library of Sweden trained its KB-Whisper models on over 50,000 hours of Swedish and reports word error rates of 5.4% on FLEURS, 4.1% on Common Voice and 5.2% on NST for its largest model, against 7.8%, 9.5% and 11.3% for whisper-large-v3. All three test sets are read speech, so none of them says how either model handles two people talking over a phone line.
Why the Nordic languages are hard for speech recognition
Danish: a sound the spelling doesn't show
Some Danish words differ in speech only by stød, a creaky catch in the voice. The lex.dk entry on stød, written by Jørgen Rischel, gives ven (friend) and vend (turn), and man (one) and mand (man), as pairs where only the second word has stød. In writing, the pairs differ by a letter. In speech, a model has to hear the stød to choose the spelling.
Stød also varies by region. The same entry notes that Jutland puts stød in words like hammer and peber where Zealand doesn't, and that a large area south of the "stød line" speaks a regional Danish without it, a fully acceptable variety. A Danish test set recorded only in Copenhagen tests one pattern.
Compound words
Danish and Swedish write compounds as one word, and the difference is audible. The Danish Language Council's guide to one or more words contrasts en tysklærer (a teacher of German) with en tysk lærer (a teacher who is German): the compound has its strongest stress on the first part, while the two-word phrase stresses both equally. Sweden's Language Council makes the same point in its quick writing rules: you say strö socker with a small pause and strösocker without one, and compounds should almost always be written together.
Word error rate punishes a split compound twice. If the reference has tysklærer and the model writes tysk lærer, that is one substitution and one insertion, two errors for one reference word. Character error rate counts one inserted space. Neither metric knows that the two spellings mean different things, so check compounds in the error list, not only in the score.
Some compounds have two correct spellings. The same Danish guide allows both en kæmpesucces and en kæmpe succes, so a reference and a correct model output can disagree without either being wrong. Decide before scoring whether such pairs are normalized to one form.
Norwegian: dialect speech, two written standards
The NPSC paper puts the problem plainly: there is no spoken standard of Norwegian, and speakers use their dialect even in official settings. Written Norwegian, though, follows Bokmål or Nynorsk, the two official written standards, and both allow many alternative forms. The authors' example is vet and veit (know): a model that writes one when the reference has the other is counted wrong, although a reader would hardly see a mistake.
NPSC's answer is to transcribe each speaker consistently in the standard they prefer, which gives about 13% Nynorsk. A product team needs the same kind of rule, written down: which standard each speaker's reference uses, which variant forms are accepted as equal, and whether dialect words are kept or replaced with a standard equivalent.
The same choice faces anyone who transcribes Norwegian for people to read. Spirelight, our parent company, runs a transcription tool for interviews in Danish, Swedish, Norwegian and Finnish, and its page on transcribing Norwegian interviews asks students to decide before they proofread whether the transcript stays close to the dialect or is normalized to Bokmål or Nynorsk.
Finnish: long words and no fixed spelling for speech
Finnish is unrelated to the other three, and it builds long words from stems and endings. The authors of the Lahjoita puhetta corpus paper note that one long Finnish word can correspond to several English words, so a single wrong ending costs a whole word in WER, and they report CER alongside it for that reason.
Colloquial Finnish adds a second problem. Speakers shorten and merge words, and because Finnish spelling follows pronunciation closely, each pronunciation can be written as it sounds. The authors conclude there is no fixed correct transcription style. When they piloted professional transcription companies on the same hour of audio, the transcribers disagreed with each other by 13.6% to 20.5% at word level and 4.7% to 6.3% at character level. If human references differ that much, a Finnish WER means little until the spelling convention is fixed.
Swedish in two countries
Finland's constitution names Finnish and Swedish as the national languages of Finland. A model for Finnish users may need to handle Swedish as well, and Swedish recorded in Sweden is not a test of Swedish spoken in Finland. The Language Bank of Finland's parliamentary collection has a Finland Swedish corpus too, of 6.4 hours, so for that variety expect to record your own test set.
English inside the sentence
English turns up inside Nordic words, not only between them. The Danish Language Council writes compounds with English parts as one word, such as webredaktør (web editor) and teambuildingkursus (team-building course), while Sweden's Language Council observes that English often writes its compounds apart. A model that writes team building kursus has made three errors on one reference word. Tag English words in your references and score them as their own slice; our article on code-switching in speech recognition covers why models trip on mixed-language speech.
How to evaluate a model on Danish, Swedish, Norwegian and Finnish
- Test each language on conversation. FLEURS and Common Voice make good smoke tests, but they are read sentences. For Danish, CoRal's conversational test split has 8.74 hours from 24 speakers. For Norwegian, the NPSC authors tested on NB Tale's spontaneous module, whose 229 native speakers in 12 dialect groups give between 7 and 8 hours. For Swedish and Finnish, plan to build one; our guide to ASR test sets covers how.
- Keep test speakers out of training and slice by region. Report the Jutland, Trøndelag or Finland Swedish result next to the average. A pooled score is dominated by whichever group has the most audio.
- Report CER beside WER. Compounds and Finnish word endings make WER swing on small differences, and CER shows whether the model was close. Our word error rate guide explains both metrics.
- Write the normalization before you score. Fix how compounds with two correct spellings are written, which Norwegian variant forms count as equal, which written standard each speaker's reference uses, and how numbers, casing and punctuation are handled. Apply the same rules to the reference and the model output.
- Match the transcription style. The National Library of Sweden publishes KB-Whisper in three styles, from condensed subtitle text to a stricter, more verbatim version, and a model scored against a reference in a different style loses points for style alone. Our article on verbatim and clean verbatim transcription shows how large that effect can be.
- Set the language and track English. Fix the decoding language for each test file rather than relying on detection, and report the code-switched slice separately so that a change in English handling can't hide in the average.
The four Nordic datasets in the SpeechData.ai catalog
SpeechData.ai has one conversational dataset for each of the four languages. Hours, speakers and prices come from each dataset page; the last column is hours times the list price.
| Dataset | Recorded in | Hours | Speakers | Price per hour | Full dataset at list price |
|---|---|---|---|---|---|
| Danish | Denmark | 500 | 50 | $95 | $47,500 |
| Swedish | Sweden | 500 | 50 | $95 | $47,500 |
| Norwegian | Norway | 500 | 50 | $95 | $47,500 |
| Finnish | Finland | 500 | 50 | $95 | $47,500 |
Together that is 2,000 hours from 200 native speakers. All four are spontaneous two-person conversations recorded in-country, with one channel per speaker, time-aligned and human-verified transcripts, and age, gender, region and recording device metadata for every speaker. Every speaker signed explicit commercial-use consent, and the full consent chain is provided for legal review. The licence covers commercial use, including training and evaluating ASR, TTS and other voice AI models.
Every speaker's region is in the metadata, so ask about the regional mix when you request a sample and compare it with your users. A worked example: a voice agent for Denmark and Norway with 250 hours of each language costs 500 × $95, which is $47,500 at list price. Volume and multi-language discounts are available on request, and the other European languages are compared on the European speech data page.
When to license and when to collect
Use open data where its terms and speech type fit. Common Voice, FLEURS and NST are good for bootstrapping a model and for a first benchmark, and parliamentary corpora add unscripted speech at scale. Check each licence against your use: CoRal, for example, can't be used for speech synthesis, and Donate Speech can't be used commercially without a paid licence.
License conversation when your product hears people talking. If the model will meet spontaneous Danish, Swedish, Norwegian or Finnish in calls, meetings or voice agents, the published results above point to conversational training data, not more read speech. A licensed dataset gives you that conversation in volume, with consent and speaker metadata you can slice by region.
Collect to spec when the catalog doesn't match your users. That includes dialect quotas the catalog's regional mix doesn't cover, Swedish speakers in Finland, a domain such as customer service or in-car commands, a particular device or room, or a language outside the catalog such as Icelandic. Through custom collection, native speakers are recruited to quotas for age, gender, region and accent and recorded on the devices and in the environments you specify, with feasibility and a quote within one business day.
A practical plan uses all three: open data to start, conversational data for fine-tuning, and a held-out conversational test set for each language and region. To check the catalog against your model, request audio and transcript samples of the Nordic datasets with your company email.
