Models and languages
The five speech models with their sizes and how fast and how accurate each one is and the download and the Remove button and the language setting with auto detect.
Dictate turns speech into text with a Whisper model that runs on your own computer. No audio is sent anywhere. You pick the model and the language in Settings under the Speech group. This guide compares the five models and explains the download and the Remove button and the language setting.
The five models
Every model is a trade between size and speed and accuracy. A bigger model hears more words right and takes longer to answer and takes more space on disk. The sizes below are what the download dialog shows before the download starts.
| Model | Download | In plain words |
|---|---|---|
| tiny | 72.0 MB | The fastest. Answers almost at once. Misses more words than the others. Good for short commands and for a slow computer. |
| base | 138.5 MB | Still fast. A step up in accuracy from tiny. |
| small | 461.1 MB | The default. A good balance for dictating sentences and emails in everyday language. About one second for a few seconds of speech on an Apple silicon laptop. |
| medium | 1457.1 MB | More accurate than small on names and accents and noisy rooms. Noticeably slower. |
| large-v3-turbo | 1542.9 MB | The most accurate of the five. Slower than small and best on a fast computer when accuracy matters more than speed. |
Every model is multilingual. There is no English only variant to choose.
Download a model
Nothing is bundled with the app so every model is downloaded once from huggingface.co into the Dictate folder on your computer. The first dictation offers the download of the chosen model. You can also open Settings and the Speech group and click Download next to the model. The state line next to the model reads Downloaded or Not downloaded.
The dialog names the model and its size before anything is fetched. Click Download and a progress bar shows the bytes. Cancel stops the download and removes the partial files so a half file never counts as a model. When the download stops on its own because the connection dropped Dictate shows a message and removes the partial files. Start it again from Settings when you are back online.
Once a model is on disk Dictate loads it from that folder with no network call. It stays loaded while Dictate runs so the second and every later dictation answers without a warm up.
Switch and remove models
Pick another model in the Speech group. When the new model is on disk Dictate unloads the old one and loads the new one. When it is not on disk the download dialog opens. The model list is disabled while a dictation is listening or transcribing so a switch never lands in the middle of one.
The Remove button deletes the files of the selected model from the Dictate folder and frees the disk space. The model reads Not downloaded afterwards and the next dictation offers the download again.
Languages
The language setting in the Speech group reads Auto detect by default. With auto detect the model listens to the first seconds and decides the language itself. This works well when you dictate in one language per sentence.
Pick one fixed language from the list when you always dictate in that language or when auto detect lands on the wrong one for a short phrase. The list holds the 100 languages Whisper knows shown by their English names such as German and Spanish and Arabic and Japanese. A fixed language is also a little faster because the model skips the detection step.
Accuracy tips
Speak at a normal pace with the microphone within arm's reach. A headset or the built in microphone of a laptop both work. Pause for a moment before you release the talk key so the last word is complete. Add the names and terms you use often to the vocabulary field so the model expects them. The Vocabulary and replacements guide explains the field and the replacements table.