AI notes in Danish: which tools support it and how to test

Short answer first: several tools now list Danish, but two of the best known do not. Otter transcribes six languages and Google Meet's Take notes for me eight, and Danish is on neither list. Word's Transcribe lists Danish (Denmark), and so do Microsoft Teams intelligent recap, Zoom's meeting summaries, Fathom, Good Tape and Notibo. Apple lists Danish for dictation but not for transcribing recordings in Voice Memos. A language on a list tells you a tool accepts Danish. Whether it gets your numbers, names and "ikke" right is something you test, and the script below gives you a small, repeatable starting point.

This page is about turning Danish speech into text and notes: lectures, meetings and interviews. The language claims were reviewed against vendor documentation on 11 October 2026. There are no accuracy scores here for any tool, because a score measured on someone else's audio says little about your lecturer in a full auditorium.

Why is Danish harder for speech recognition than English?

Many tools transcribe speech before generating notes from the text. Training data and processing methods vary, so there is no basis for assuming that every model is mainly trained on English. A transcription error can affect the notes, and note generation can introduce separate mistakes or omissions. Check both outputs.

What the Whisper paper shows

OpenAI's Whisper paper is one of the few that publishes its training data language by language, which makes it a useful illustration of the problem.

• 680,000 hours of audio, most of it English. The Whisper paper says 117,000 of those hours cover 96 languages other than English, and a further 125,000 hours are translation data.
• 473 hours of Danish. The paper's appendix lists 473 hours of Danish speech recognition audio against 438,218 hours of English. Swedish has 2,119 hours and Dutch 2,077.
• Less data, more errors. The authors report a squared correlation of 0.83 between the logarithms of training-data amount and word error rate on Fleurs. Their fitted relationship estimates a halving of word error rate for a 16-fold increase in data; it is not a universal causal rule.
• Some errors are not mishearing. The paper describes failure modes in long recordings: getting stuck in repeat loops, dropping the first or last words of a segment, and complete hallucination, a transcript unrelated to the audio.
• Results differ by language. OpenAI's Whisper README says plainly that performance varies widely depending on the language.

Not every tool is built on Whisper, and the paper describes models from 2022, so read the numbers as an illustration of the pattern, not as a measure of any product you can buy. The practical lesson is to check the language and recording conditions you need, rather than infer a modern product's quality from this older training set.

Which tools list Danish?

The table records what each vendor's own page says. "Listed" means Danish is named on that page. It says nothing about quality.

If your Microsoft 365 account includes Transcribe and permits the intended data use, Word is worth checking before buying another subscription. Microsoft's Transcribe page lists Danish (Denmark) among more than 80 locales and gives 300 minutes of uploaded audio a month on a Microsoft 365 subscription, or 30,000 minutes with a Copilot licence. It runs in Word for the web in Edge or Chrome, takes .wav, .mp4, .m4a and .mp3 files, and returns a transcript, not notes. That covers a lecture or your own notes. For interviews and other recordings with other people's personal data, your programme's rules come first. SDU's library, for example, writes that the transcription function in Word does not comply with the GDPR rules, points SDU students to Transcriber on UCloud instead, and tells them to check their programme's policy on AI and data before they transcribe.

Apple is the odd one out. Apple's feature availability page lists Danish for dictation and for transcribing audio messages in Messages, but not for audio transcription in Notes, and the Voice Memos guide names ten languages without Danish. So an iPhone takes Danish dictation but will not transcribe a Danish recording in Voice Memos. Good Tape, made in Copenhagen, names Danish on its languages page and claims more than 100 languages in all. Among meeting apps, Fathom and Fireflies list Danish too, and their prices and limits are in the meeting notes app comparison.

What goes wrong in Danish speech?

When you test a tool, include details where a small change would matter: quantities, names, negations and subject terms. The examples below are possible failure cases to check, not measured error rates.

• Numbers and times. Danish says 95 as femoghalvfems, with the unit first and a tens word built on twenties, and "halv tre" means half past two, not half past three. A transcript can put digits in the wrong order or move a meeting by an hour.
• Names. Use names such as Søren, Aabenraa and Brønshøj to check spelling, including æ, ø and å. A model's total Danish training hours do not tell us how often any particular name appeared.
• Negations. "Ikke" is short and easy to swallow in fast speech. A transcript that drops it turns "vi har ikke besluttet det" into "vi har besluttet det", and the notes then report the opposite of the meeting.
• Compound words. Danish joins nouns into one word: dækningsbidrag, forsyningskæde, eksamensforberedelse. A tool can split them in two or spell the joint wrong, which also breaks searching your notes later.
• English terms in a Danish sentence. "Vi skal have feedback på modellen inden deadline" mixes two languages in one breath. Danish with English subject terms is different from switching the whole meeting language. Test the actual phrases; Otter's documented French-English exception does not establish Danish support.
• Silence. The Whisper paper names hallucination as a failure mode. Put 20 seconds of silence in your test recording: unexpected words there need checking against the actual audio, including any background speech.

How do you run a small Danish comparison?

A recording of only your own voice reading fictional details avoids involving classmates in the initial test and gives you a known script. Still check the service's terms, storage and sharing settings. The names, budget and meeting in this script are invented. Read it at your normal pace and use the same file and settings where possible for each tool. Avoid accidentally capturing other people in the room.

The script

1. Start the recorder on your phone, or the tool itself, in the room you normally use.
2. Read lines 1 to 7 at your normal speed, without slowing down for the hard words.
3. Stay silent for 20 seconds, then read line 9.
4. Run the same file through two tools.
5. Paste each transcript next to the script and mark every line that differs.

How to score it

Distinguish harmless formatting differences from errors that change meaning. Record the kind of error: a wrong number, a misspelt name, a missing ikke, a split compound, a mangled English term, or words in the silence. If the tool also writes notes, check whether the missing ikke or the wrong time made it into the summary. That is the error that costs you later.

Nine lines can reveal particular mistakes; they are not enough to rank performance for an entire term. The one mistake a tally can miss is a vanished negation in a sentence that still reads well, so read lines 4 and 7 slowly. Once you have permission to record a real lecture or meeting, repeat the test on that, because one clear voice in a quiet room is the easy case.

Is a Danish transcript enough to study from?

A transcript can help you inspect wording and revisit an explanation, but it is not a complete revision method. Transcription styles vary, and automatic output may omit or change words. Organise the explanation, check uncertainties and practise answering course questions. Errors can pass from the transcript into notes, but they do not all necessarily do so. Transcription is not the same as notes goes through the difference in detail.

It also changes what you compare. Word's Transcribe and Good Tape focus on transcription; compatible Apple devices also offer summaries. Teams recap, Zoom and Fathom provide meeting summaries and other AI features. If you want study notes and flashcards from a Danish lecture, you need a tool that goes that far, or you do the second step yourself.

Where does your Danish audio go?

Ask two questions, because they have separate answers. Where are the files stored, and where is the speech turned into text? A tool can store your recordings in the EU and still send the audio to a speech service in another country.

• International transfers need an applicable legal route. The European Commission's page on standard contractual clauses describes contractual safeguards for certain transfers outside the EU/EEA. They do not on their own establish a lawful basis for collecting a recording or make every use compliant. Other transfer routes, such as adequacy decisions, can also apply.
• Storage and processing are different answers. A privacy page that says "stored in the EU" has answered the first question only. Look for the list of processors and where they run.
• Other people's voices raise the stakes. A lecture, a meeting or an interview contains other people's personal data, so their permission and your institution's rules come before any upload, and for interview data some programmes name the tools you may use.
• A vague answer is a no. If a vendor's pages do not say where speech-to-text runs, ask in writing before you upload anything you would not post publicly.

Your own voice reading fictional details reduces the personal information involved. It does not remove the need to understand the service's handling of the recording. Resolve institutional requirements before using classmates' or interviewees' data.

What mistakes do people make when choosing?

• Trusting the dropdown. A language in a menu means the tool accepts it. Otter's own settings article notes that German and Chinese are still in beta, a reminder to read feature-specific limitations as well as the list.
• Testing on a podcast. Clean studio sound may not represent your lecture room. Test on your own script, then in the room you actually record in.
• Judging the transcript only. The transcript is the input. The notes, where a dropped ikke becomes a wrong decision, are what you will read.
• Assuming the platform handles Danish. Google Meet's notes do not, while Teams recap and Zoom's summaries do. Check the platform you actually meet on. If you are moving away from Otter for this reason, Otter.ai alternatives compares the options.
• Recording other people to run a test. Use your own voice for the test, and get permission before you record a lecture or a meeting.

Where Notibo fits

Notibo transcribes 78 languages, Danish among them, and the transcript is intended to follow the language spoken. It records through a computer or phone browser, or accepts an uploaded audio file. It returns a transcript and structured notes; recording-based flashcards and spaced repetition require Pro, including during the trial. On a phone, microphone recording needs the page kept open and screen unlocked. It also turns PDF, PPTX and DOCX files into notes and flashcards. One recording can last up to two hours within the remaining allowance. Long recordings are saved in smaller parts as they run and joined into one transcript, and media uploads are capped at 50 MB.

None of that replaces the test above. Run the script through Notibo and one other tool, and compare the lines yourself. Files are stored in the EU, on Supabase in Frankfurt, while transcription and the AI notes run on US processors under standard contractual clauses, and source audio is normally deleted after the transcript and notes are saved. Keep a permitted original if you need to listen again. The privacy policy lists the processors. Pro is free for 14 days, with 240 minutes of recording and no card, so the test costs nothing. After the trial the free plan gives 30 minutes of recording a month, and keeping Pro costs 9.99 euro a month or 88.99 euro a year; the lecture recording page shows how a recording becomes notes.