Greek: lose the accent marks and half the words fall apart

Once we gave Doubao the right Greek language code, it produced real Greek, and the meaning matched. But every vowel carrying an accent mark (the tonos) vanished entirely. κάνεις became κνεις, έχουμε became χουμε, and Μπήκε shrank to Βκη. About half the words came out broken.

Our first suspicion was that our own processing was stripping the marks. It wasn't. In the same pipeline the other four engines kept their accents intact (81 to 87 each); Doubao had zero. This is how Doubao itself behaves. It didn't invent anything or flip any meanings. It was simply too broken to serve as a reference track.

The test clips: all four on the clear side

• Dogtooth (Κυνόδοντας, 2009), two scenes: five people talking around a dinner table, full of numbers; and a security guard's visit followed by a father's medical questions, with quick back-and-forth and a long medical word for appendicitis (σκωληκοειδίτιδα).
• Alps (Άλπεις, 2011), two scenes: a rushed exchange at a hospital front desk, including the name Jude Law; and a home visit where someone reads from a fashion magazine, full of foreign names like Ferré, Guatemala and Winona Ryder.

All four come from the same director, whose style is deliberately cool and slow: one person finishes, then the next speaks. On our three-tier scale most of this is baseline. We went through more than forty films looking for a noisy scene with people talking over each other and couldn't find one with both the audio and Greek subtitles available.

Engines we tested

ElevenLabs Scribe v2, Speechmatics, AssemblyAI and Doubao (el-GR). Soniox transcripts from a separate all-engine run were compared during the review. Gemini was added in August 2026.

Red lines, engine by engine

AssemblyAI had problems on all four clips, and they carried over when we switched films:

• Flipped meaning: "Μπήκε χειρουργείο" (went into surgery) became "Βγήκε η χειρουργία" (the surgery came out). Turned around completely.
• Lost passage: in the magazine scene, the stretch with the three foreign names was simply missing.
• Loop: at the end, three words piled up on a single timestamp, repeated.
• Invented words: Jude Law became a name that doesn't exist, and in "water the yucca", the yucca turned into a made-up word, Ιούκαντ.

Speechmatics mostly dropped articles and prepositions, the safest kind of mistake. But it made two unsafe ones: in the magazine scene, "interview with Winona Ryder" became the English "Coming soon"; at the hospital, "another one the same" became "and the truth". It also heard "want coffee?" (Θες καφέ) as "yesterday coffee" (Χθες καφέ).

ElevenLabs: no hallucinations. It spelled the long medical word correctly, got Jude Law right, and got all three foreign names in the magazine right (Ferré was missing one letter).

What we chose, and why

Primary: ElevenLabs Scribe v2. Lowest error rate on three of the four clips; on the fourth it lost only on how it wrote numbers. Reading line by line, it made almost no mistakes.

Reference 1: Speechmatics. A steady second on all four clips, close to the primary. Most of its errors were omissions, and the dangerous ones were isolated.

Reference 2: Soniox. Better than AssemblyAI on all four clips, with accent marks intact.

Reference 3: Gemini. In the August round it had no red lines on the dinner scene and the lowest error rate there, 11.9%. On the home-visit scene it made one slip, turning "schnitzel" (Σνίτσελ) into "usually" (Συνήθως).

AssemblyAI isn't in the standing lineup, and Doubao is out.

Evidence against our choice

• Speechmatics is listed ahead of Soniox out of habit, not because we measured it. In the review, Soniox read at least as well as Speechmatics on three clips. With both tracks running, the order doesn't matter for now. If we ever have to keep only one, we'll need to compare them again.
• Speechmatics' "Coming soon" is English that came out of nowhere, which is not a minor thing. It stays because it happened once across four clips, and because Doubao can't be used for Greek, so there are few engines to take its place.
• Among the original four engines, Speechmatics had the lowest error rate on the dinner scene.

Still open

• No noisy, overlapping-speech clips. This is the biggest gap. AssemblyAI stumbled even on clips this clear; how any engine holds up in a noisy room, we don't know.
• No interview material. We found a Greek production with excellent subtitles for this, but the only "full" copy online had the wrong audio attached, so we had to drop it.