Voice dictation for English speakers in Israel
Speech recognition hears an English sentence with a Hebrew word in it and guesses an English word that sounds close. Hebrish is a fine-tune of NVIDIA's Parakeet that writes the Hebrew word itself, in Latin letters, and runs on your own phone or computer.
I want to go to the Micollette today to buy my groceries.
I want to go to the makolet today to buy my groceries.
Words stock models trip over, and Hebrish writes as one consistent spelling:
Free. Runs entirely on your device. The model downloads once (about 660 MB) and the app checks now and then for a newer version.
Install steps and other platforms: download page. No install: try it in the browser.
Hebrish runs inside Parlance, a free dictation app for Android, Ubuntu and Windows. Pick the English-Hebrew code-switching model in it; Parlance lets you switch back to stock Parakeet in one tap.
Parlance is a voice keyboard. Pick the code-switching model, tap the mic in HeliBoard, talk, and the text lands where you're typing. Try it now opens a notepad with a mic button, so you can test it before setting anything up.
Works with HeliBoard. Early build. Setup guide
A tray app. Press a hotkey or a foot pedal anywhere, talk, and the text is pasted at your cursor. A Transcribe tab handles longer recordings and audio files.
Install: sudo apt install ./parlance_0.1.0-1_amd64.deb
The same tray app as a portable zip: unzip it and run Parlance.exe.
Early build, not yet tried on real hardware.
pip install parakeet-mlx.Parakeet's vocabulary has no Hebrew letters, so Hebrish doesn't output Hebrew script. It learns one romanized spelling for each Hebrew word and writes that, punctuated and cased like the rest of the sentence.
It was trained on read-aloud sentences recorded on a phone: English with one or two Hebrew words dropped in, about 300 minutes so far. About 30% of the training audio is plain English, because fine-tuning on code-switched speech alone is known to wreck a model's English.
The training recordings are one speaker's: an Irish-accented English speaker living in Jerusalem. It works best for voices like that. If yours is very different, stock Parakeet may suit you better; Parlance lets you switch between them in one tap.
Version 3, scored on held-out recordings that were never trained on. Some of the Hebrew terms in the test set never appear in training at all.
| Test | Stock Parakeet | Hebrish v3 |
|---|---|---|
| Code-switched English, word error rate | 10.9% | 5.3% |
| Hebrew terms seen in training, written correctly | 29% | 66% |
| Hebrew terms never seen in training, written correctly | 27% | 48% |
| Plain English, same speaker, word error rate | 3.5% | 2.9% |
| LibriSpeech test-clean, word error rate | 1.2% | 1.0% |
| Long recordings (~3 min), word error rate | 9.0% | 5.0% |
Lower word error rate is better. English got slightly better, not worse. Versions 4 and 5 are published and not yet benchmarked. On the separate public benchmark, v3 leads on Hebrew term recall at 67.4%, against 45.9% for Whisper Large V3 and 32.6% for stock Parakeet. Every version and how it was trained: model card. Full history: project tracker.
The model, the recordings and the benchmark are all public. The base model is CC-BY-4.0, and so is the fine-tune.