Choose a model¶
A model is what turns your voice into text. Each one downloads once, about 660 MB, and then works offline.
| Model | What it is for | Start here if |
|---|---|---|
| Sabich | English with Hebrew words and tech vocabulary, updated as new recordings come in | You want the recommended model: it is the default on Android, Ubuntu and Windows |
| Shakshuka | Streaming: words appear while you speak | You want to see the words as you talk. On computers only for now, and it lags behind Sabich too |
| Parakeet (stock) | NVIDIA's release, unchanged: standard English, without the Hebrew words | You want plain English only |
The Models screen in each app lists these with estimated speed and accuracy, and marks the recommended one. On a computer it can list more variants, and Help me choose… asks about your hardware and recommends one. The same guide is at hebrish.app/models/choose. Measured results are on Benchmarks.
All of the fine-tuned models write Hebrew words in Latin letters, because that's what they're for: English text with Hebrew words in it. They don't output Hebrew script.
The models are trained on one speaker's voice, so they work best for that voice and may do less well for yours. More voices are welcome: see Get involved.
Commercial use¶
The models are free for personal use, research, education and small non-profits. Commercial use needs a paid licence, and so do health funds and public bodies. Licensing explains the rest.
Next: how to download and switch models on desktop or Android.