German: when Viennese dialect made the steadiest engine switch to English

In a drinking-table confrontation from the Austrian film Hundstage, several men needle each other in thick Viennese dialect. On this scene AssemblyAI suddenly switched to English, producing lines like "You just often want to sing... I will the luz dance like if I was a toad," none of which is in the audio. On the same scene, Doubao stayed in German throughout and was the most coherent of the group.

On the other three clips, all standard German, AssemblyAI was the cleanest engine we tested. Within a single language, a dialect can turn the ranking upside down.

The test clips

Four film scenes, three German and one Austrian, roughly two to three minutes each:

• Knockin' on Heaven's Door (1997): two leads talking fast in a diner, interrupted by a gangster's phone call.
• Der Untergang (2004): a bunker war council, five or more officers cutting each other off.
• Sophie Scholl (2005): a courtroom interrogation with a crowd shouting over it. This is the only clip where the subtitles match the audio word for word, so its WER is meaningful.
• Hundstage (2001): Viennese dialect, bullying at a drinking table. The usable reference covers only 1 minute 42 seconds.

Several other films were on our list, but we couldn't get them past YouTube age restrictions, or the "full film" turned out to be an English-language recap. Good sources for hard German film dialogue are scarce.

On the three tiers from the first post, Knockin' and Sophie Scholl sit near the middle tier, and Der Untergang and Hundstage are in the hardest. The basic tier isn't covered.

The candidates

ElevenLabs, AssemblyAI, Doubao, Speechmatics (two models), Soniox, Gladia, Deepgram, FunASR, and later Gemini.

The six red lines, engine by engine

Invented words and sentences. In Sophie Scholl, the crowd shouts "Frechheit" (outrageous). Three engines each invented a different word for it: Soniox wrote "Brecher als sowas Brecher Schwein oder Schwein," Deepgram wrote "Wecker" (alarm clock), and Gladia wrote "Lecker als sowas Lecker Lecker" (tasty). In Knockin', Soniox heard "Frankie ist harmlos" (Frankie's harmless) as "ein Harvard" and added a sentence of its own. In Der Untergang, Deepgram turned "kulturlosen Bestien" into a string of made-up words.

Language drift. That's the scene at the top. When we re-checked, AssemblyAI had switched to English three separate times on the Austrian clip.

Flipped meaning. On the Austrian clip, Speechmatics turned "Dir vergeht die Lust, wenn du tot bist" (you'll lose the urge once you're dead) into "Ich will dich los wenn du tot bist," changing both the subject and the meaning, and added an insult, "du Schwein."

Dropped passages. Deepgram skipped about 30 seconds of dialogue in Knockin'. On the Austrian clip it returned 25 words, missing roughly 90%.

Repetition. Soniox showed mild repetition on the Austrian clip, with one word appearing four times in a row. We have no such record for the others on these four clips.

What we chose

Main track: ElevenLabs Scribe v2. It never switched language, never invented a sentence, and made between zero and three substitutions per clip.

Reference tracks, in priority order:

1. AssemblyAI (universal-2): cleanest on the three standard-German clips, with zero errors on the Der Untergang war council.
2. Doubao: stayed in German on all four clips, with no invented sentences and no large gaps. On Austrian dialect it was steadier than AssemblyAI.
3. Gemini: added on August 26, 2026. When we tested it across 26 languages, German was in the group where it could go straight in. On Hundstage it had no red lines and didn't switch to English where AssemblyAI did.

Evidence against our choice

• Doubao bends meanings. It wrote "Frankie ist harmlos" (Frankie's harmless) as "hart vor," pushing the meaning the other way. "Ja, no problem" became "ja nur Brot" (just bread), and "Acht, Acht" (eight, eight) became the number 88.
• The main track changed a meaning too. In Der Untergang, ElevenLabs turned "In einem Krieg wie diesem" (in a war like this one) into "In einem Kollektiv dieser Art" (in a collective of this kind).
• Gemini invented a sentence in German. In Der Untergang it wrote a line nobody says, "Flucht oder Berlin wird ein gewaltiger Rückschlag," and turned "Are you mad?" into "He's mad."
• The one trustworthy WER winner didn't make the cut. Speechmatics had the lowest WER on Sophie Scholl. On standard German it really is safe; its failures are mostly omissions. But on the Austrian dialect it changed both the subject and the meaning of a line.

Open questions and limits

• We have one Austrian clip, with only 1 minute 42 seconds of usable reference. Swiss German wasn't tested at all.
• All four clips are from films. None is a real interview, and none is in the basic tier.
• In the few seconds of crowd shouting in Sophie Scholl, three engines each invented a word, while AssemblyAI and Speechmatics left the gap empty. An empty gap is safer than an invented word, but it shows those seconds are still hard for most engines.