Hebrish
Experimental Hebrish is an early-stage proof of concept, trained on a small code-switching dataset in one speaker's voice. First results are promising. Growing the training data is a long-term effort, and the aim is for every version to get better. Roadmap

Try it

Try Hebrish in your browser

Say a sentence with a Hebrew word in it. The Hebrish model writes it down, right here on your computer. Nothing to install, and your audio never leaves your device.

Large download The model is about 660 MB. Your browser downloads it once from Hugging Face, then keeps it, so the next visit starts without downloading again. It works best on a desktop or laptop in Chrome, Edge or Firefox. Once loaded, the model takes about 2.5 GB of memory, so close other heavy tabs first.

On a phone the download is heavy and the page may run out of memory. The Android app is the better way to try it there.

Everything runs on your device. Your recordings and files are not uploaded anywhere.

1. Load the model

Checking whether the model is already saved in this browser…

2. Say something

Press Record, read one of the sentences below (or your own), then press Stop. Up to 30 seconds.

Sentences to read aloud

These lines come from the dataset. The Hebrew word should come out in Latin letters, spelt as shown. The Hebrew above it is only there for reference; the model never writes Hebrew script.

  • I need to go to the iriyaעירייה to pay my arnonaארנונה bill.
  • Did you see where I put my teudat zehutתעודת זהות?
  • Do you want ambaעמבה on your sabichסביח or just tahini?
  • The office was a complete balaganבלגן after the move.

More words it knows: the lexicon.

What to expect

  • One speaker's voice. The model was fine-tuned on recordings of one person. It may do worse on your voice and accent.
  • An early proof of concept. It will get some words wrong, Hebrew and English. It knows the words in the lexicon, not every Hebrew word.
  • Short clips work best. It is built for dictation: a sentence or two at a time.

Want to compare it with stock Parakeet, or skip the download? The Hugging Face demo runs both models on a server and shows the two transcripts side by side. For everyday dictation, get the app.

How this page works

The page runs the int8 ONNX build of the model (version 5) with parakeet.js and ONNX Runtime Web, on your processor. The model files are stored in your browser's own storage; Remove the saved model deletes them. More about the model: the model card.