Fluen Studio does not rely on a single speech-to-text engine. We work with five providers, test them side by side, and send each language to the engine that handles it best. Here is how that choice is made, what we look for, and when it makes sense to pick an engine yourself.
The engines only produce the raw transcript: the words and when they were spoken. Turning it into readable, well-timed subtitles is a separate step that Fluen does in-house, whichever engine was used. We call it subtitle segmentation.
How Auto works
Every workspace starts with the transcription engine set to Auto (Recommended). Auto follows our own test results, not the providers' marketing claims, and changes engine as soon as a better one shows up in our tests.
- 1You uploada video or audio file, in any of 50+ languages
- 2Auto picksthe engine that tested best for that language
- 3Fluen checksthe result, and retries with another engine if it fails or invents text
- 4Fluen splitsthe transcript into readable, well-timed subtitles
What we test
A low error count is only part of the story. Some mistakes cost far more time in subtitles than their share of the words suggests, so our benchmark looks for them on purpose.
| Check | Why it matters for subtitles |
|---|---|
| Word accuracy | Every wrong word is a correction you have to make |
| Word timings | Subtitles are timed from each word, so a word with no length can throw off a whole line |
| Language detection | Files left on automatic detection must be recognised correctly from the first seconds |
| Writing system | Cantonese and Traditional Chinese must come out in Traditional characters |
| Dialects | Some engines rewrite colloquial speech into the standard form of the language |
| Mixed-language audio | Some engines translate the second language instead of transcribing it |
| Invented text | Over silence or music, some engines repeat themselves or make up words |
Word timings are a good example of a difference you would never see in a transcript, but would feel in your subtitles:
Share of transcribed words with a start time but no length. Lower is better.
What our latest benchmark showed
We tested on two kinds of material: clean read speech in 28 languages, and real long-form recordings in 34 languages that are publicly available online, such as talks and interviews.
- No single engine won everywhere. Five different engines and models came first in at least one language.
- Real recordings are harder. Clean read speech flatters every engine, and the gaps between them shrink on real-world audio. That is why we test on real recordings, not only on read speech.
- The details matter. The engine that looks best on average is not always the right one for a given language, which is exactly what Auto is for.
Each engine has its own strengths and weak spots. The fastest engine, which also has perfectly clean timings, is the one that most often gets the language wrong:
Short clips of 7 to 15 seconds in 28 languages, with the language left on automatic detection. Higher is better.
Median processing time per minute of audio, across all test files. Lower is better.
The engines
| Engine | Strengths |
|---|---|
| ElevenLabs Scribe v2 | Very accurate on clean and real-world audio, reliable on long recordings, good with names and terms (especially with a glossary) |
| Speechmatics Enhanced | Very accurate, strong on Chinese varieties, writes Cantonese and Traditional Chinese in Traditional characters |
| Deepgram Nova-3 | Precise word timings, follows speakers who switch languages even mid-sentence |
| AssemblyAI Universal | The widest language coverage, fast and consistent |
| OpenAI Whisper | Natural, well-punctuated text that reads as if already edited |
| Deepgram Whisper | Whisper models on Deepgram's platform, suited to long recordings |
Should you pick an engine yourself?
For almost everyone, no: Auto already uses the engine that tested best for each language. If you have a reason to, here are your options:
| If you want to... | Do this |
|---|---|
| Use one engine for a whole workspace | Workspace Settings > Subtitle Defaults > Transcription engine. It applies to every new file, in every language. |
| Retry one file that came out poorly | Open its captions, click Troubleshoot and choose Try a different AI engine. No extra cost, once per track. |
| Get names and terms right | Add them to your glossary. It guides every engine. |
Speaker detection runs as its own step, so it works the same whichever engine is chosen.
Accuracy matters, readability matters more
Getting the words right is the foundation. But in subtitles, where a line breaks and how long it stays on screen matter just as much. That part is ours, and it meets the same standard whichever engine produced the transcript. Read more in our overview of subtitle splitting.