Skip to main content

Part of: 27 languages, one by one

Danish: we swapped out the cleanest engine on the most interview-like clip

· Zeyu Si

On this page

Of our five Danish test clips, the one closest to a real interview is a conversation at a rehab centre: one person asks, the other answers, and they work out goals together. On that clip, AssemblyAI produced the cleanest transcript of any engine.

We took it out of our reference engines anyway.

The test clips

Five clips, from two films by two directors:

Clip Scene Difficulty
Jagten (2012), part 1 A family confrontation that turns into a brawl: shouting, heavy background noise Extreme
Jagten, part 2 A supermarket refuses to serve a customer on Christmas Eve; staff shove and beat the customer; store noise Extreme
Adam's Apples (2005), part 1 A priest and a neo-Nazi arguing about the Book of Job and the devil. Emotional but clearly spoken Medium-high
Adam's Apples, part 2 A gang causing trouble, insults, a gunshot, several people shouting at once Extreme
Adam's Apples, part 3 A goal-setting conversation at a rehab centre. Question and answer, quick turn-taking, clear indoor audio Basic

We wanted to use the dinner scene from Festen (1998) and had the subtitles ready, but couldn't find the original Danish audio. The only full-length copy was age-gated; another turned out to be the French dub.

Engines we tested

ElevenLabs Scribe v2, AssemblyAI, Speechmatics, Soniox and Doubao, with Gemini added in August.

Doubao doesn't support Danish. What came out was English gibberish with Chinese characters mixed in, plus loops.

The red lines

Loops: at the climax of the Jagten brawl, AssemblyAI wrote "Nej det er satan" 23 times in about 28 seconds, swallowing the real dialogue. It didn't loop once in the other four clips.

Negation flips: Speechmatics turned "Det gør jeg ikke" (I don't) into "Det gør jeg" (I do), so a denial became an admission. In the supermarket, the accusation "so it's fine to hit the customers?" became "you normally have a good meeting with the customers". In Adam's Apples, "he shot me" became the command "shoot me".

Hallucinated sentences: in the brawl, ElevenLabs turned "Tror du ikke det" (don't you think so) into "Jeg dør dig", which isn't a real sentence.

Meaning-changing substitutions: AssemblyAI likes to invent sound-alike words. Koteletter (pork chops) became koaguletter, "the layer cake" became "legislation", and "why are you lying" became "why are you spying". On the rehab-centre clip, Soniox turned "cake" into "banana cake" and "my goal is to bake" into "my mum makes". On review we classed those two as mis-segmentation with an acoustic basis, not reversals.

Language drift: Doubao only.

What we chose

  • Primary engine: ElevenLabs Scribe v2. Cleanest, or joint cleanest, on all five clips, and it never looped. Its one hallucination in the brawl didn't derail the rest of the scene.
  • Reference engines: Speechmatics, Soniox, Gemini.

The bar for a new reference engine is zero red lines across all six categories, plus a top-two overall ranking among the non-primary engines. Soniox cleared it on review and replaced AssemblyAI. The overall ranking came out Speechmatics first, Soniox second, AssemblyAI third.

Speechmatics stays in first place because its word accuracy was slightly better than Soniox's. It never loops and its output holds its structure.

Gemini was clean on the rehab-centre clip, and had at least two clean conversational clips overall.

Evidence against our choices

There's plenty for Danish.

  • AssemblyAI looped once, in an extreme physical fight, the kind of scene a real interview almost never contains. On the most interview-like clip it was the cleanest engine. If you weigh "how it does on the clips closest to production" above "how many red lines it crossed", keeping AssemblyAI is a defensible call.
  • Soniox changed meanings on the quiet rehab-centre clip too. Those confident errors weren't forced by noise.
  • When we tested Gemini in August, we logged more Soniox problems on the same clips: in the brawl it twice heard "lying" as "promising" and added an invented sentence at the end; in the shooting scene, "he shot me" twice became "can you shoot me?". That conflicts with its zero-red-line rating, and we haven't re-assessed it yet.
  • Speechmatics flipped meaning several times across the five clips (at most once per clip), yet it holds the top reference slot. We removed AssemblyAI over red lines without holding Speechmatics to the same standard. That isn't consistent.
  • Gemini turned "you should go" into "I should go" in the shooting scene, and changed a pronoun in the brawl.

What's still open

  • Soniox's rating needs to be checked again against the August notes.
  • All five clips come from two films, and none is a real research interview. If we can get Festen's original audio, it would add a dinner-table overlap scene.

Our Danish transcription page

Written by Zeyu Si