Indonesian: someone screams for help, the engine writes 'thank you'

One of our Indonesian test scenes is a cult laying siege to a house. Through the screaming, someone shouts "Bu, tolong selamatkan anak-anakku, bu!" ("Mum, please save my children!"). Our primary engine, ElevenLabs, got that line. At the same moment, AssemblyAI wrote "Terima kasih", which means "thank you".

Gladia went further. It produced "Jangan lupa like share dan subscribe", "don't forget to like, share and subscribe", the sign-off from a YouTuber's video.

Both point the same way. When the audio gets too messy to make out, some engines fill the gap with a familiar phrase.

The test clips

All four clips come from one film, Satan's Slaves (Pengabdi Setan, 2017), scored against its Indonesian subtitles for the deaf and hard of hearing. We used a single film because every other Indonesian "full movie" we found on YouTube was fake: a review video, a different film under the same title, or a re-cut whose timing didn't match the subtitles.

The four scenes each stress something different:

• Leafing through a magazine: string music underneath, two people firing lines back and forth.
• Family dinner: clattering dishes, younger siblings messing about, constant overlap.
• Asking for help: two people calmly telling a family story in full sentences.
• The siege: panicked screaming, thunder, someone battering the door.

On our three-level difficulty scale the first two fall between middle and top, the siege is top level, and the help scene is close to easy, the only clear clip in the set.

Engines we tried

ElevenLabs Scribe v2, AssemblyAI, Doubao and Speechmatics in the first round, with FunASR, Gladia, Deepgram, Soniox and Qwen run alongside for the record. Gemini was tested separately.

Where the engines crossed red lines

Invented sentences. The two in the opening: AssemblyAI's "thank you" and Gladia's "like, share and subscribe". AssemblyAI invented nothing in the first three clips. Only the siege pushed it into one.

Inserted words. In the siege someone says "kalau kita bersatu" ("if we stand together"), and none of the four engines got it. ElevenLabs wrote "gang up on", AssemblyAI and Speechmatics wrote "hand over", and Doubao wrote "kalau kita nggak serahkan" ("if we don't hand over"), adding a negative nobody said. At dinner, Speechmatics also stuck two extra words, "hati kau" ("your heart"), on the end of a line.

Repetition loops. In the siege, people keep shouting the name Bondy. ElevenLabs, Soniox and Doubao each wrote the name five to nine times. Gemini wrote it as Bodoh ("stupid") ten times over.

Dropped passages. AssemblyAI lost the most in the siege, leaving out long stretches of the screaming.

Language drift. Doubao's first two clips came back as a mix of Chinese and English gibberish, because we hadn't sent the Indonesian language code. With id-ID set, its transcript of the help scene was almost word for word the same as the primary's.

None of the engines flipped the meaning of a whole sentence in any of the four clips.

What we chose

Primary: ElevenLabs Scribe v2. Lowest WER on the first two clips, next to no meaning errors on the clear one, the most content recovered from the siege, and the only engine that caught the cry for help.

References, in order: Soniox, Doubao, Gemini. All three run on every Indonesian file.

• Soniox: replaced AssemblyAI. Review confirmed no red-line errors, and it ranked first or second among the non-primary engines. Its two questionable spots both came in extremely noisy passages where every engine failed, and what it wrote matched sounds in the audio, so they count as mishearing, not invention.
• Doubao: once the language code was fixed, no invented sentences and no reversals in four clips. It comes from a very different technical background than Soniox, so the two don't tend to fail in the same place.
• Gemini: on the clear help scene, it matched the primary and Soniox almost word for word.

AssemblyAI became a backup. The invented "thank you" plus the screaming it dropped put it closer to our red lines than Soniox, judged by the same standard.

Evidence against our choice

• Soniox has a habit. When it can't make something out, it tends to write a fluent sentence instead of leaving obvious garbage, and a fluent wrong sentence is the hardest kind for a reader to spot. An early automated scan also flagged signs of over-generation and repetition in its Indonesian output. The configuration notes that, in the first month of production, fluent hard-to-hear lines that only Soniox produced should be spot-checked.
• Doubao's extra negative in the "stand together" line is exactly the kind of error that does the most damage when transcripts are merged.
• AssemblyAI was clean in the first three clips. The case for replacing it rests mostly on one scene.

Open problems

• All four clips come from one film, with the same cast and the same recording conditions. Other films and other accents might give a different answer.
• The siege scores are unusable, so the ranking there comes entirely from reading line by line.
• We have no real interview recordings.