QuietSteno

Home / Guides / Choosing an accuracy setting

Choosing an accuracy setting

The three settings are three sizes of the same speech model. Larger ones make fewer mistakes and take longer. Here is how to pick.

Choose Balanced for most recordings on a laptop or desktop. Choose Fast on a phone, on an older computer, or when you only need a rough draft of clear speech. Choose Most accurate for recordings with accents, background noise, several speakers, many names, or languages other than English, when you can wait longer.

What the settings are

QuietSteno uses Whisper, an open-source speech recognition model published by OpenAI. Whisper comes in several sizes. A bigger model has more parameters, recognises speech more reliably and needs more memory and time.

SettingWhisper modelDownload with a graphics chipDownload withoutRelative speed
Fasttiny114 MB39 MBQuickest
Balancedbase197 MB73 MBMedium
Most accuratesmall558 MB237 MBSlowest

Each model is downloaded the first time you choose that setting and then kept in your browser’s cache, so you only wait once per setting.

Why there are two download sizes

Browsers that support WebGPU can run the model on your graphics chip, which is far faster. That path uses a higher-precision version of the model, so the files are larger. Browsers without WebGPU run a compressed version on the processor: a smaller download, but slower processing. QuietSteno detects which your browser supports and chooses automatically. The status line tells you whether it is using the graphics chip or the processor.

Which setting for which recording

  • One person, clear microphone, English: Balanced. Fast is often good enough.
  • Podcast or interview with two or three people: Balanced, or Most accurate if you will quote from it.
  • Meeting or lecture recorded from the room: Most accurate. Distance and echo are hard on small models.
  • Strong accents or non-native speakers: Most accurate.
  • Languages other than English: Most accurate. Small models are noticeably weaker outside English.
  • Translating speech into English: Most accurate.
  • On a phone: Fast, and keep recordings short.

If the result is poor

  1. Check the language setting. The wrong language is the most common cause of a nonsense transcript.
  2. Move up one setting and use “Run again with these settings”.
  3. Look at the recording. If you struggle to follow it yourself, no setting will rescue it. A cleaner copy of the audio helps more than a bigger model.

Known limits of every setting

  • Silence and music can cause a speech model to invent a few words. QuietSteno skips sections that are completely silent, but stray text can still appear over music or noise.
  • Repeated phrases occasionally appear when the audio is unclear. Delete the duplicates when you edit.
  • Speakers are not identified. Add names by hand.
  • Very long files need a lot of memory. If the browser runs out, split the recording.

Transcribe a file now

Questions and answers

Which accuracy setting should I use?

Balanced for most recordings on a computer, Fast on phones or for quick drafts, and Most accurate for difficult audio, non-English speech and translation.

Why is the download so large?

The download is the speech model itself, which runs on your device instead of on a server. It is fetched once per setting and cached by the browser.

Does Most accurate take much longer?

Yes, it is the slowest of the three and needs the most memory. On a computer with a capable graphics chip it is still practical for long recordings; on a phone it may not run.

Can I delete the downloaded models?

Yes. Clear the site data for QuietSteno in your browser’s settings. The models will be downloaded again the next time you transcribe.

Related

Last updated 13 November 2026.