Swedish: when "apologize!" became "I apologize"

In the Swedish film Ondskan (Evil, 2003) there's a boarding-school bullying scene. The bully yells "Be om ursäkt" at the other boy: apologize. ElevenLabs, Speechmatics, Soniox and Deepgram all transcribed it as "Jag ber om ursäkt", I apologize. The person demanding the apology became the one giving it.

Only AssemblyAI got it right. That line is part of why AssemblyAI is still in our Swedish lineup.

The test clips

Four clips, all from film and TV. We don't publish the audio.

The first two exist to make engines fail. The last two are closer to a real interview.

Engines we tested

ElevenLabs Scribe v2, AssemblyAI, Speechmatics, Soniox, Deepgram, Qwen, FunASR and Doubao. We added Gemini in August 2026.

Swedish isn't on Doubao's published list of 25 supported languages. Feed it Swedish and you get English back, sometimes looping: "I don't know," fifteen times in a row. It was out for this language.

The red lines

Negation and stance flips were the most common problem in Swedish. The opening line is one. ElevenLabs also dropped an inte in Tillsammans, so "don't look at them, don't listen to them" came out as "don't look at them, listen to them." Speechmatics turned "Tycker du" (do you think) into "Jag tycker" (I think), and changed "the clumsiest" into "the most dishonest", which turns a remark about clumsiness into an attack on someone's character.

Hallucinated sentences — we first pinned three on AssemblyAI in the two group scenes: lines that weren't in the subtitles. On review, Soniox and Qwen heard the same words in the same places. The subtitles had left out real dialogue. Since then, when several engines agree against the reference, we check the reference first.

Loops: Qwen tended to repeat itself in Swedish; our automated scan counted one word repeated 35 times in a short stretch. Speechmatics turned the crowd chanting "Kryp!" into twelve "Hi"s. Strictly that's a substitution, but the meaning is gone either way.

Language drift only showed up with Doubao. We didn't see a clear case of dropped passages or inserted content from any engine on these clips.

What we chose

• Primary engine: ElevenLabs Scribe v2. The most complete and faithful across the four clips, with no loops. Both of its flips happened in the group scenes; on the two interview-style clips it had none.
• Reference engines: AssemblyAI, Soniox, Gemini.

AssemblyAI: it was the only engine to get the apology line right, and it was clean on both interview-style clips.

Soniox replaced Speechmatics. Speechmatics had been on record as "no flips". When we went back through the clips, that didn't hold up: Speechmatics was among the engines that got the apology line wrong, and the other two changes above were its. Soniox had no red line on these four clips that was unique to it. At each point where Speechmatics lost marks, Soniox either got it right or left a gap rather than guess.

Gemini joined in August. Zero red lines on both interview-style clips. On Morgonstudion it produced 730 words, the most complete transcript of the group, with book titles and names intact.

Evidence against our choices

• ElevenLabs flipped meaning twice too. It works as the primary engine because the reference engines can catch it at points like these.
• Gemini turned "you broke up the family" into "I broke up the family" in Tillsammans.
• An early automated scan found Soniox producing 1.8 times as many words as the reference in Swedish, with repetition. A line-by-line review later found no red line unique to Soniox, but we haven't forgotten the signal.
• On Babel, the cleanest clip, Speechmatics scored 32.7% WER, in the same lowest band as Qwen's 32.5%. ElevenLabs scored 44.0%. We still dropped Speechmatics, because what we count is how often the meaning goes wrong, and on that measure it did worse.

What's still open

Four clips: two film group scenes, two TV segments. None of them is an actual research interview, and we didn't pick any for regional accents. And cases like the opening line, where several engines make the same mistake and only one gets it right, can't be fixed by a majority vote. So on negation and reversed meaning we don't go with the majority: without hard evidence, the line is flagged for a person to listen to.