Large downloadThe model is about 660 MB. Your browser downloads it once, then keeps it. Best on a desktop or laptop in Chrome, Edge or Firefox; once loaded it uses about 2.5 GB of memory.
On a phone, the Hebrish app is the better way to try it: the same model, offline, as a keyboard.
Open the full page to try it
Here, inside Hugging Face's frame, your browser won't let the model use more than one processor core, and it would freeze while it starts. The full page runs it properly.
1. Load the model
Checking whether the model is already saved in this browser…
2. Say something
Press Record, read a sentence below (or your own), then press Stop. Up to 30 seconds.
What Sabich heard
Sentences to read aloud
Hebrew words should come out in Latin letters, spelt as shown. The Hebrew script is only a reference; the model never writes it. Numbers come out as digits.
- I need to go to the iriyaעירייה to pay my arnonaארנונה bill.
- Did you see where I put my teudat zehutתעודת זהות?
- Can you check the GitHub Actions logs before we deploy to Vercel?
- The invoice came to 1,250 shekels, due on the 15th.
Every Hebrew word it was taught: the lexicon.
Sabich or Jachnun?
Sabich is the main Hebrish model and the one the apps recommend: code-switching plus tech vocabulary, brand names, numbers and everyday dictation. Jachnun is the code-switching-only model; you can try Jachnun here. Both are fine-tunes of NVIDIA Parakeet TDT 0.6B v2, trained on one speaker's voice. Results against stock Parakeet are on the benchmarks page.
If you speak English with Hebrew in it, send us a few recordings and a future version learns your voice too.
How this page works
It runs an int8 ONNX build of Sabich v2.3 (model card, access by request), downloaded from Hebrish's own CDN, with parakeet.js and ONNX Runtime Web on your processor. The files are kept in your browser's storage; Remove the saved model deletes them.