Hebrish · היברישVoice typing for English with Hebrew words mixed in, the way English speakers in Israel talk: "pick up bamba from the makolet", "I have an appointment at Misrad Hapnim tomorrow". The models write each Hebrew word in one consistent romanized spelling (arnona, kupat holim, teudat zehut), never Hebrew script, and run on your own device. A personal project by Daniel Rosehill in Jerusalem.
These are the models kept up to date. If you are not sure, pick Sabich Rolling.
The main Hebrish model: English with Hebrew words mixed in, plus tech vocabulary, numbers and Israeli names. It is retrained as new recordings come in, and it is what the Hebrish apps run. Fine-tuned from NVIDIA Parakeet TDT 0.6B v2.
For voice commands: the kind of thing you say to an AI agent or to your computer. Two sizes: Falafel on Parakeet TDT 0.6B v2 and Falafel 110M on Parakeet TDT-CTC 110M.
General English and tech vocabulary, without the Hebrew code-switching. A separate project by the same author, licensed CC-BY-NC-4.0.
Also available: Shakshuka for live, streaming transcription (Nemotron streaming base), and Meorav, a text-to-speech model that speaks Hebrish.
The easiest way to use Hebrish. The apps download the models from Hebrish's own CDN, so no Hugging Face account is needed.
curl -fsSL https://cdn.hebrish.app/apt/setup.sh | sudo sh) or the .deb.These stay on Hugging Face for reference, reproducibility, citations and existing installs, but they are no longer updated and are now behind the models above. Each card says so at the top.
| Repository | What it is | Use instead |
|---|---|---|
| Hebrish/Sabich | Sabich v1.0 to v2.4, the released versions before the rolling line (+ sherpa-onnx, ONNX, GGUF, MLX builds) | Sabich Rolling |
| Hebrish/Jachnun | The first stable release (v1.0, 2026-10-19), an experiment no longer maintained (+ GGUF, ONNX, sherpa-onnx, MLX) | Sabich Rolling |
| Hebrish/Sabich-Qwen3-ASR-1.7B | Proof of concept: Sabich's data on a Qwen3-ASR base, trained once (+ GGUF) | Sabich Rolling |
| Hebrish/Sabich-Canary-Qwen-2.5B | Sabich's data on NVIDIA Canary-Qwen 2.5B (GPU, server use) | Sabich Rolling |
| Hebrish/Sabich-Parakeet-110M | Small Sabich on Parakeet TDT-CTC 110M (+ sherpa-onnx, GGUF) | Sabich Rolling, or Falafel 110M for commands |
| Hebrish/Sabich-asr4all-S, -M | Small Sabich builds for wearables and small boards | Sabich Rolling |
| Hebrish/Whisper-Hebrish-Small-ACFT | Whisper small.en fine-tune for FUTO Keyboard and whisper.cpp | Sabich Rolling |
| danielrosehill/Whisper-Hebrish | Whisper large-v3-turbo fine-tune, the first Hebrish experiment | Sabich Rolling |
| danielrosehill/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching | Proof-of-concept checkpoints PoC-1 to PoC-5, kept for their DOI | Sabich Rolling |
| Hebrish/Parakeet-TDT-0.6B-v2-Tech-And-Accent | Boxty: English and tech without Hebrew, Fritter's predecessor | Fritter |
Every model with its formats, sizes and published results: Hebrish Models and hebrish.app/models.
The models are fine-tuned on recorded sentences of English with Hebrew words in them, written in a canonical romanization, plus a tech vocabulary corpus and general English. The training data so far is one speaker's voice (an Irish-accented English speaker living in Jerusalem); recordings from more speakers are the next step, and contributions are welcome. It transcribes English with Hebrew words in it, not full Hebrew sentences.
Most model repositories are gated, but approval is automatic: fill in the short form on the repository page and accept the terms, and you have access straight away. Use is free for personal, research, education and non-commercial work under the Hebrish terms of use; models from 2026-10-25 on are CC-BY-NC-4.0. Commercial use needs a commercial licence: see licensing or write to access@hebrish.app.
Updates: the newsletter (at most one email a month), the blog and the podcast. Feedback and feature requests: hebrish.app/feedback.