Fluid Speech Pro, in your browser

Since 20 November 2026, on-device transcription in Cuecaps runs Fluid Speech Pro by FluidVoice, the most accurate open speech model under 1B parameters (#2 among open models on the Open ASR Leaderboard, 6.06% average word error). It runs inside your browser tab, so it's free, unlimited, and your video never uploads. Every word comes back with its own timing, punctuation and capitals, ready for word-by-word captions.

On this page

How to use it

  1. Open cuecaps.com/studio and drop in a video or audio file. No account needed.
  2. Click Find words and choose On this device.
  3. Leave Speech model on Most accurate and pick the language.
  4. The first time, the model downloads once (about 800 MB) and stays in your browser for next time.
  5. Every word lands on the transcript and the timeline with its timing. Pick a caption style and export.

What changed

  • Accuracy: Fluid Speech Pro replaces Whisper base as the default. It has 0.63 billion parameters to Whisper base's 74 million, and is the most accurate open model of its size.
  • Punctuation and capitals come from the model itself, so there are fewer commas to fix by hand.
  • Every word is kept, um and uh included, so Fast edit can still cut the filler words and retakes.
  • It's still free and unlimited on every plan, and it never uses cloud minutes.

Languages

Fluid Speech Pro covers English and 16 more of the studio's languages: Spanish, Portuguese, French, German, Italian, Dutch, Danish, Swedish, Finnish, Polish, Russian, Ukrainian, Czech, Greek, Hungarian and Romanian.

For every other language in the studio, including Japanese, Chinese, Korean, Arabic, Hindi and Turkish, on-device transcription uses Whisper base automatically, and the Find words dialog says so. Cloud transcription covers 90+ languages.

The smaller download

Prefer a quick first download? Choose Smaller download under Speech model in Find words to use Whisper base instead, under 300 MB. Cuecaps remembers the choice. If Fluid Speech Pro can't run on a device, for example one short on memory, the studio switches to Whisper base by itself.

To free the space later, open the command menu (Ctrl K) and choose Clear the downloaded speech model.

How it runs

  • The 8-bit ONNX version of the model, run by parakeet.js on ONNX Runtime Web inside a worker in your browser tab.
  • Your audio is never sent anywhere: the model and its libraries download once from Hugging Face and jsDelivr, then stay in your browser's cache.
  • On a desktop computer, a 20-second clip transcribes in about 7 seconds. Speed depends on the computer or phone.
  • Fluid Speech Pro by FluidVoice is built on NVIDIA Parakeet TDT 0.6B v3 and released under CC BY 4.0.

Questions

Which speech model does Cuecaps use on my device?

Fluid Speech Pro by FluidVoice, the most accurate open speech model under 1B parameters, for English and 16 more languages. Whisper base covers the studio's other languages and is the smaller download you can pick instead.

Is Fluid Speech Pro free in Cuecaps?

Yes. On-device transcription is free and unlimited on every plan, with or without an account, and it never uses cloud minutes.

Why is the download about 800 MB?

The model has 0.63 billion parameters, even in its 8-bit version. It downloads once and stays in your browser, so later runs start in a few seconds. Choose Smaller download for Whisper base, under 300 MB.

Does it upload my video or audio?

No. The model runs in your browser tab. Only the model files download, once; your video and audio stay on your computer or phone.

Is it more accurate than cloud transcription?

On clear speech it comes close. For music under the voice, strong accents, several speakers or languages outside its 17, cloud transcription (ElevenLabs Scribe v3) is still the better choice, and it takes your own names and terms.

Try it on your video

You can also drop a file anywhere on this page. Or tryTry the demo clip

See Creator, $9 a month

Caption your video