Romanian: "Are you hungry?" became "I'm starving"

The opening of a Romanian film: two people talking back and forth, very little background noise, no crowd. Of our four Romanian clips, this is the clearest.

One of them asks: "Ți-e foame?" Are you hungry?

Speechmatics wrote "mor de foame": I'm starving. A question became a statement, and it's now coming from the other person. In the same minute it turned "listen, please" into "I heard it too", "I took the flowers" into "I gave the flowers", and replaced the "tort" (cake) in "meringue cake" with a word that doesn't fit, satisfăcător ("satisfactory").

When an engine does this on the easiest clip in the set, the problem is the engine, not the recording.

The test clips: three films, four scenes

• Filantropica (2002): a crowd outside a nightclub, heavy noise, people shouting over each other. Hardest tier.
• The Death of Mr. Lăzărescu (Moartea domnului Lăzărescu, 2005): an emergency room, with doctors, nurses, patients and phone calls crossing over each other, plus medical terms like bariu (barium). Mid-to-high tier.
• 4 Months, 3 Weeks and 2 Days (4 luni, 3 săptămâni și 2 zile, 2007), two scenes: a birthday dinner with a whole table talking across each other, constant interruptions and lots of prices (hardest tier); and the opening argument between two people described above (baseline tier).

Romanian sources are scarce. Films with audio available lack subtitles, and films with subtitles lack full audio, so two scenes had to come from the same film.

Engines we tested

First round: ElevenLabs Scribe v2, AssemblyAI, Speechmatics and Doubao (ro-RO). FunASR, iFlytek and Qwen were run for reference only. Soniox transcripts came from a separate all-engine run and were brought into the review in early July. Gemini was added in August 2026.

Red lines, engine by engine

Speechmatics had problems on every clip. Besides the ones above: on the street, "we're going to Cult Club" became "we're going to bed", and a simple "Where?" grew into "Where is Lake Sloop", a place nobody mentioned. In the emergency room it added a sentence nobody said, roughly "the doctor's on the phone".

AssemblyAI didn't loop, but it invents words in crowded scenes. "Cult Club" twice became a non-word, "Kulfland". At the dinner it twice turned "dragă" (dear) into the curse "dracu". On the clear clip it inserted a negation, so "Go on! I'll tell you…" became "Go on, I won't tell…". It also garbled a name, broke the verb at the start of a sentence, and left out whole stretches.

ElevenLabs: no flips on any clip. Its mistakes were dropping "Cult", and twice hearing "sutimi" (hundredths, as in cents) as the similar-sounding "sutieni" (bras).

Doubao: no hallucinations, no flips, and every price at the dinner right. Its mistakes were sound-alike swaps, such as ceaiul (the tea) for ceva (something) in "can I bring you something from downstairs?"

Soniox: no fabricated passages; its transcripts were about the same length as ElevenLabs' and Doubao's, not padded. It got the name Popescu and the conditional "dacă-mi vii" (if you come to me) right, exactly where AssemblyAI garbled and Speechmatics flipped. It blurred two sound-alike spots on the clear clip.

What we chose, and why

Primary: ElevenLabs Scribe v2. Lowest error rate on the clear clip, no flips on any clip.

Reference 1: Soniox. Among the reference candidates it had the lowest error rate on all four clips, and it read clean line by line. It replaced AssemblyAI. AssemblyAI's mistakes are mostly odd words that stand out, so it went for being weaker than Soniox, not for being dangerous.

Reference 2: Doubao. No hallucinations, no flips. Soniox and Gemini are both large-model transcribers; Doubao takes a different approach, which makes it a useful counterweight.

Reference 3: Gemini. See the next section.

Evidence against our choice

• The original conclusion was AssemblyAI plus Doubao. It only changed after Soniox's transcripts were pulled out and reviewed. The body of that original report still states the old conclusion, with a correction added at the top.
• Speechmatics was the only engine to get the medical term "bariu" right both times. Soniox wrote "bari", one letter short.
• The primary engine gets things wrong too: ElevenLabs heard "cents" as "bras", twice.
• Gemini had no clean film clip in Romanian. In the emergency room it turned "bariu" into "bani" (money) and repeated itself at the start. At the dinner it made four or five errors, dropping the core of one sentence and turning "whisky?" into "juice?". Then we ran a five-minute real interview podcast. No hallucinations, no flips, no loops; two minor slips ("I" as "he", and a muddled idiom), and three places where it was more accurate than ElevenLabs. That podcast had no human reference transcript; we compared Gemini against ElevenLabs only.

Still open

• Four scenes from three films, and none of them a research interview.
• Only the emergency-room clip has specialist vocabulary, and the engines differed a lot on it (see "bariu" above). Findings there are not yet solid.