Because OmniaScribe runs the speech model on your own device, you get a choice that upload services make for you: how big a model to use. A bigger model catches more words, names and accents. A smaller one finishes sooner and works on modest hardware. Neither is better in every case.
This guide explains the three sizes (Fast, Standard and High accuracy), the difference between English-only and multilingual versions, and a simple way to decide. It does not quote speed or accuracy figures, because they depend on your device and your recording; it shows you how to find out on your own.
The three sizes
Every size does the same job, listening to speech and writing it down. They differ in how much the model has “learned” and how much work your processor must do for each second of sound.
- Fast: the quickest and the smallest download. It suits clear speech, short files and older or budget phones. It is also the one most likely to stumble on accents, background noise and uncommon words.
- Standard: the balance most laptops and phones handle well, and the one the home page marks as Recommended. It is a sensible default for interviews, lectures and meetings.
- High accuracy: catches more words, accents and names. It is best on a laptop or desktop, and it takes noticeably longer than the smaller sizes on the same file.
In practice the differences show up in the details: a surname, a technical term, a mumbled phrase or someone speaking quickly with an accent. For a clean recording of one person speaking clearly, the three sizes often read almost alike.
English-only or multilingual?
Each size comes in two versions. The English-only version is trained for English alone. The multilingual version understands many languages, and can detect which one is spoken.
If your recordings are always in English, the English-only version is the better pick. In the accuracy stars each model carries, every English-only version scores the same as or higher than the multilingual one of the same size for English. It is the one to choose for American, British or Canadian English.
Choose multilingual when you need another language, a mix of languages or automatic detection. It is also required for translation: only a multilingual model can translate speech to English, so the English-only versions cannot, as the translate audio to English page explains. OmniaScribe supports many languages in total, and the languages page shows an honest rating for each model size.
Try it on your own file
What your device can handle
A model has to fit in memory and finish in a reasonable time. A recent laptop copes with any size. An older phone may struggle with the largest, and you will feel it as a longer wait. The Models screen looks at your device and recommends a size, and it times a downloaded model on your device, so the labels reflect your own hardware instead of a guess.
Length limits are tied to size too. On a computer, Fast handles up to 3 h, Standard handles up to 2 h, High accuracy handles up to 1 h. On phones it's half of that. So a two-hour recording is a Fast or Standard job on a computer, and a long phone recording may need to be split. Files expected to take longer than about 45 min ask before they start, and the message suggests a lighter model.
A simple way to decide
- Start with Standard (English) if you only work in English, or Standard (Multilingual) if you do not.
- Transcribe a minute or two of a real recording and read it against the audio. Count how many corrections you needed.
- If it is mostly right and quick, stay. If the errors are in names and hard words you care about, try High accuracy on the same clip.
- If it is slow or your device struggles, try Fast on the same clip and see whether the loss is acceptable.
- Remember the answer for that kind of recording. A podcast with a good microphone and a noisy meeting room may need different choices.
Using the same short clip for every test is the key. Comparing models on different recordings tells you more about the recordings than the models.
Suggestions by job
- Quick voice notes and memos: Fast or Standard is usually enough. The speech to text page records and shows a live draft while you speak.
- Lectures and long classes: Standard keeps the wait sensible, and High accuracy is worth it only for the hardest material. See the lecture transcription page for tips.
- Interviews with names and quotes: Standard, or High accuracy when quotes will be published and your device can take the time.
- Noisy rooms and accents: step up a size before you give up on a recording.
- Translation to English: a multilingual model of Standard size or larger.
Downloads, storage and switching
A model is downloaded once and then stored in your browser, so later transcriptions need no connection. You can download more than one and switch between them any time. The Models screen shows what is installed and lets you delete a model you no longer need to free up space; you can download it again any time.
The speech model is a large download, so on mobile data it is worth waiting for Wi-Fi. A download that is interrupted can pick up where it left off, and every file is checked after it arrives, so a bad download is caught instead of used.
No model replaces proofreading. Even High accuracy makes mistakes, and every size is best treated as a very fast first draft. The choice is about how much cleanup that draft needs, and how long you want to wait for it.