Part of: 27 languages, one by one
Czech: the two-letter prefix that decided our engine lineup
· Zeyu Si
On this page
In Czech, "not" is often nothing more than ne- attached to the front of a verb.
Our hardest Czech clip is a family scene with everyone around the table talking at once. At one key negation, Speechmatics dropped the ne- and the sentence came out meaning the opposite. Soniox kept it. That clip is where we changed our Czech reference engines.
The test clips
Four clips from two films:
| Clip | Scene | Difficulty |
|---|---|---|
| Pelíšky (1999), part 1 | The family climbs on the table to act out how tall a bear is: fast exchanges, interruptions, swearing, background noise | Extreme |
| Pelíšky, part 2 | An argument at dinner. Clearly spoken but tense, with dense vocabulary | Medium-high |
| Kolja (1996), part 1 | A police interview, one or two people questioning one, occasional interruptions | Basic |
| Kolja, part 2 | The 1989 Velvet Revolution: radio news playing over a crowd shouting in the street | Extreme |
Engines we tested
ElevenLabs Scribe v2, AssemblyAI, Speechmatics, Soniox, Deepgram, Gladia, FunASR and Doubao. Gemini was added in August.
Doubao doesn't support Czech. The four clips came back either empty or as a mix of English and Chinese nonsense, including loops like "No no no".
The red lines
Negation flips: the one at the top, from Speechmatics.
Dropped passages: AssemblyAI lost two whole passages in the hardest family clip. On the revolution clip it produced 78 words in total, against 154 from Speechmatics: about half.
Loops: Doubao, yes. ElevenLabs wrote the crowd's "Už je to tady!" (It's here!) six times on the revolution clip, where the subtitles have it twice. The crowd really was chanting it over and over, so we didn't count it as a loop, but we noted it.
Inserted content: on the revolution clip, Gemini wrote an English stage direction, "(Crowd noise and shouting)", into the transcript, and turned "bells ringing" into zabili (they killed).
Language drift: Doubao only.
Hallucinated sentences: in the original four-clip evaluation, none of ElevenLabs, AssemblyAI or Speechmatics invented a whole sentence. Their errors were mostly sound-alikes, and a few changed the meaning. Speechmatics heard má slon in "the elephant is three metres seventy!" as máslo (butter), and "I'd never swallow it" as "I'd never recognise it".
What we chose
- Primary engine: ElevenLabs Scribe v2. The most faithful on three of the four clips. The dinner argument came out almost word for word.
- Reference engines: Soniox, Speechmatics, Gemini.
Soniox: zero red lines across the four clips. On the hardest family clip it tied with ElevenLabs for best, and it was the only engine to spell one character's name, Brašule, correctly. It kept the key negation.
Speechmatics stays, with one rule attached: if Speechmatics is the only engine with a particular reading of whether a sentence is negative, the merge step doesn't go with it.
AssemblyAI moved to backup because of the dropped passages above.
Gemini had zero red lines on the clear police interview.
Evidence against our choices
- Our first four-clip report concluded that AssemblyAI and Speechmatics only ever dropped words and never reversed meaning, so three engines were plenty. Re-running the hardest clip overturned that. Our first read was too optimistic.
- On the revolution clip, Speechmatics caught the most of the radio broadcast and was the best of the three engines compared at the time; ElevenLabs came last. When we tested Gemini in August, the notes for that clip say ElevenLabs dropped a large part of the opening and ranked last. Our primary engine isn't reliable on noisy broadcast-plus-crowd audio.
- Gemini has only one conversational clip to its name in Czech. On the other, it inserted a stage direction and changed a meaning.
What's still open
- All four clips come from two films. None is a real interview.
- Gemini needs more conversational material.
- Speechmatics has dropped a negation, and our fix is a rule not to trust it on its own. That patches the merge step; the engine's weakness is still there.
Written by Zeyu Si