About the project
About Hebrish
A small, open project to make dictation work for the way English speakers in Israel actually talk.
Why it exists
English speakers in Israel say Hebrew words all day. You pay the arnona, renew your teudat zehut, pop out to the makolet. Speech recognition hears those words and writes whatever English words sound closest. Every dictated message then needs fixing by hand.
Hebrish fixes that one problem. It is English dictation that writes the Hebrew words correctly, in Latin letters, on your own device.
What it is
Hebrish is a fine-tune of NVIDIA Parakeet TDT 0.6B v2, a small, fast English speech model. Parakeet v2 was chosen over v3 on purpose: in Daniel's own dictation v2 was more accurate for English, and v3 trades some English accuracy for 25 more languages.
It was trained on read-aloud sentences: English with one or two Hebrew words dropped in, mostly recorded on a phone (a OnePlus Nord 3 5G), with a small number of desktop-mic recordings. About 30% of the training audio is plain English, because fine-tuning on code-switched speech alone is known to damage a model's English. Each training run takes well under an hour on one NVIDIA L40S GPU and costs under a dollar.
You use it through Parlance, a free dictation app for Android, Ubuntu and Windows that lets you pick this model or stock Parakeet.
One voice, for now
Every training recording is one speaker's: an Irish-accented English speaker living in Jerusalem. It works best for voices like that, and gains on other voices are untested.
Everything is open
The model and the dataset are CC-BY-4.0. The code is on GitHub.
- Model.nemo, GGUF, ONNX, sherpa-onnx
- DatasetThe code-switched recordings, with transcripts
- BenchmarkA frozen test set no model trained on
- LeaderboardModels side by side on that test set
- AppsParlance source code
- LexiconThe romanized spelling of each Hebrew term
- TrackerWhere the project stands, every version
- Write-upHow v3 was trained and what changed
Who makes it
Hebrish is a personal project by Daniel Rosehill. Daniel was born in Ireland and lives in Jerusalem, and his is the voice the model was trained on.
Most of his writing starts as voice, so he cares about accuracy on real voice notes and fast inference on a phone. Alongside Hebrish he builds other English-Hebrew code-switching datasets and models: Hebrish is the English-with-Hebrew-words that English speakers in Israel actually use, and off-the-shelf speech recognition mangles it.
- Website: danielrosehill.com
- Blog: blog.danielrosehill.com
- GitHub: github.com/danielrosehill
- Hugging Face: huggingface.co/danielrosehill
Get involved
The model gets better with more words and more voices. Four ways to help:
- Report a bug: a word it wrote wrong, or an app that misbehaved.
- Suggest a word: a Hebrew word you use in English that it should know.
- Send recordings: your voice reading sentences. This is the only way it will learn voices other than one.
- Something else: testing, apps, ideas.
Credits
- NVIDIA Parakeet TDT 0.6B v2, the base model (CC-BY-4.0). Hebrish is not made or endorsed by NVIDIA.
- NVIDIA NeMo, used to train every version.
- HeliBoard, the open-source Android keyboard Parlance plugs into.
- transcribe.cpp, used for the GGUF builds.
- sherpa-onnx, which runs the model inside the apps.
- LibriSpeech, replayed during training to keep the model's English.