Questions and answers
FAQ
What Hebrish does, what it doesn't, and how to help it get better.
Using it
What is Hebrish?
Voice dictation for English speakers in Israel. You speak English with Hebrew words in it, and it writes the Hebrew words correctly, in Latin letters: makolet, not Micollette.
It is a fine-tune of NVIDIA's Parakeet TDT 0.6B v2 speech model. It runs on your own phone or computer, inside the free Hebrish app. You can also try it in the browser first.
What does "Hebrish" mean?
It's the mix of English and Hebrew that English speakers in Israel actually use. You speak English, and drop in Hebrew words where they fit: arnona, teudat zehut, kupat holim.
Linguists call switching languages mid-sentence code-switching. That's the technical name for what this model handles.
Can I dictate whole sentences in Hebrew?
No. Hebrish is an English model that knows Hebrew words. It expects English sentences with Hebrew words in them. Speak a full sentence of Hebrew and you will get English-looking guesses, the same as with any English model.
For Hebrew-only dictation, use a Hebrew speech model.
Will it work with my accent?
It works best for voices like the training voice, an Irish-accented English speaker in Jerusalem, and recordings from more speakers are the next step.
Want it to learn your voice? Contribute recordings. And if stock Parakeet does better for you, switch back in one tap.
Which platforms does it run on?
Through the Hebrish app: an Android voice keyboard that works with HeliBoard, and tray apps for Ubuntu and Windows. The Ubuntu build is tested on Ubuntu 26.04 (amd64). The Windows build is early and not yet tried on real hardware. Links are on the download page; on Android, the setup checklist walks through the permissions.
On a Mac with Apple Silicon there is an MLX build of Jachnun v1.0 (and one of Sabich v2.3); run it with parakeet-mlx --fp32. Any other app that loads a Parakeet GGUF file can use the GGUF builds, for example Offline Voice Input on Android or Handy on Linux.
Is there a Mac or iPhone app?
Not yet; both are planned. On iPhone, Apple does not let custom keyboards use the microphone and limits their memory to far less than a speech model needs, so Hebrish can't be a voice keyboard there the way it is on Android. It would be a separate app, installed through TestFlight, which needs a paid Apple developer membership. Details: Hebrish on iPhone: what it would take.
On a Mac with Apple Silicon, developers can already run the model from the command line: download page.
What phone do I need?
Android 9 or newer, on an arm64 phone.
The Hebrish app downloads the model it needs (about 660 MB). While you use it, the model takes about 700 MB of memory, and the app frees that when you stop. If you use another app with GGUF files instead, Jachnun's GGUF build has a Q8_0 file (730 MB), or Q4_K_M (475 MB) for phones with 4 GB of RAM or less. Sabich v2.3 and Shakshuka v2.0 have only an F16 GGUF (1.24 GB), meant for desktop GPUs: on a phone, use them through the Hebrish app. Speed on phones has not been measured yet.
Which model should I pick, Sabich or Jachnun?
Sabich is the recommended model in the apps and the default. It is updated on a rolling basis: most updates continue from the previous weights, and v4.0 (2026-11-10) was retrained from scratch as a new baseline. Sabich was trained on the same English-with-Hebrew recordings as Jachnun, plus tech vocabulary, and learning the tech words did not cost it accuracy on Hebrew words: on the same 422 held-out recordings (2026-10-20), Sabich v1.0 got 5.3% of words wrong and Jachnun v1.0 5.4%, and it was a little better on tech dictation too (1.5% against 1.7%). Those are v1.0's numbers; on the newer test set v2, Sabich v2.3 (2026-10-30) gets 4.96% of words wrong, against 5.85% for Jachnun v1.0; the rolling Sabich in the apps has no published evaluation numbers. Hebrish now has three active models, Sabich for transcription, Shakshuka for streaming and Falafel for voice commands (its first release is training now); Jachnun stays available as an experimental test and is not in active development.
Both stay in the model picker. If one suits your speech better, switch in one tap. For other devices (a computer with a GPU, a wearable) or live captions, Which model should I use? picks one of the other variants.
How do I switch back to stock Parakeet?
In the Hebrish app, open the model picker and choose stock Parakeet instead of Sabich. It's one tap, and you can switch back the same way.
How it works
How does speech-to-text work?
It works on your device, on your own phone or computer: your audio is not uploaded to transcribe it. That means it also works with no internet connection, which matters most in an emergency.
The app needs the internet once, to download the model (about 660 MB). After that it checks now and then for a newer version of the model.
How was Hebrish created?
Hebrish is a fine-tune of NVIDIA Parakeet, a cutting-edge speech recognition (ASR) model efficient enough to run on your own hardware.
What is a fine-tune?
Adapting a base model with extra training so that it gets to know niche vocabulary, such as Hebrew words. It keeps what it already knows and learns the words it used to get wrong.
Why does normal dictation get Hebrew words wrong?
Speech models write the words they were trained on. An English model has never heard makolet, so it writes the English-looking words that sound closest. Stock Parakeet wrote "Micollette".
Hebrish was trained on recordings of English sentences with Hebrew words in them, so it has heard the Hebrew words and learned to write them.
Why Latin letters and not Hebrew script?
Parakeet v2's vocabulary has no Hebrew characters, so the model can't write Hebrew script. It writes each Hebrew word in Latin letters instead.
That also keeps a sentence in one direction. Mixing right-to-left Hebrew into left-to-right English often scrambles word order in messages and documents.
Which spelling does it use?
One consistent spelling per word, taken from a public lexicon of 1,164 terms. Every training transcript follows it, so the model always writes, for example, kupat holim, not a different spelling each time.
If you prefer another romanization, the model will still tend to write the lexicon's one.
How accurate is it?
The latest measured release is Sabich v2.3; the Sabich in the apps has had rolling updates since, not scored yet. On public benchmark test set v2 (2026-10-30: 129 English-with-Hebrew recordings it was never trained on):
- Word error rate on English with Hebrew words: 13.72% for stock Parakeet, 4.96% for Sabich v2.3.
- Hebrew words written correctly: 19.9% stock, 67.0% Sabich v2.3.
- Hebrew words not in the training data, written correctly: 16.1% stock, 59.9% Sabich v2.3.
- Plain English, word error rate: 2.73% stock, 2.42% Sabich v2.3.
On this test set v2.3 is slightly below Sabich v2.2 (4.91% and 70.1%), while on held-out recordings from its own dataset it is ahead; the two test sets disagree, and both differences are small.
On the older test set v1, Jachnun v1.0 wrote 73.3% of Hebrew terms correctly, against 45.9% for Whisper Large V3. Shakshuka v2.1 (2026-11-05) is not benchmarked yet. Full results are on the benchmarks page.
What's the difference between the versions?
Mostly the amount of training audio. Each version is trained from stock Parakeet on more recordings: PoC-3 on 196 minutes of code-switched speech, Jachnun v1.0 on 339 minutes, Sabich v2.3 on 522.8 minutes (1,296 minutes of training audio in all, with the tech recordings and English replay). The first five checkpoints, published as v1 to v5, were the proof of concept and are now called PoC-1 to PoC-5. Jachnun v1.0 was the first full release after the proof of concept. Since Release v2 (2026-10-22) models are retrained together as one numbered release: the current versions are Sabich v2.3 (a Sabich-only update of 2026-10-30), Shakshuka v2.0 and Meorav v2.1, and Jachnun stays at v1.0.
What changed in each release is on the changelog. How each one was trained is on the model page.
Is there a streaming model?
Yes: Shakshuka, built on NVIDIA's Nemotron Speech Streaming model and trained on the same data as Sabich. The current version is Shakshuka v2.0 (2026-10-22), a choice for dictation in the Hebrish desktop apps (0.7.0 and later; Sabich is the default); Android does not offer it yet. It is not benchmarked yet.
Its proof of concept, Shakshuka PoC-2, scored offline on the public benchmark got 62.2% of Hebrew terms right, behind Jachnun v1.0's 73.3%, and streaming at 560 ms chunks lost nothing measurable against offline (5.65% against 5.81% word error rate).
Privacy & cost
Are you training on my transcripts?
No. Inference happens on your device: your audio and transcripts aren't uploaded, and there's no telemetry. Your transcript history stays on your device, and you can turn it off entirely. You have full control over it. Details: privacy policy.
Is it free? Is it open?
Yes to both. The apps are free, and their source code is public. The models are free for personal, research and other non-commercial use (new versions are CC-BY-NC-4.0; see the licence of each version), and the training dataset is CC-BY-4.0.
For developers
Is there an API?
No public API. If you are an inference provider, hosting the models and charging for inference needs a commercial licence, and Daniel is looking into inference providers.
To run your own, Hebrish Server is a self-hosted web page and OpenAI-compatible API. The desktop app also serves its own local API to programs on your computer.
Can I use Hebrish in my app?
Yes, in line with the licensing terms. It depends on the app. If it is free and for the public good, such as a free Torah learning app, no licence is required, and Daniel would appreciate hearing from you.
The models are licensed CC BY-NC 4.0: give credit, and don't sell them. If you charge for your app, you need a commercial licence; these are not available yet: Licensing.
Getting involved
It wrote a word wrong. How do I report it?
Use the bug report form. Say what you said, what it wrote, and which app and platform you used.
Can I suggest a Hebrew word it should know?
Yes. Suggest a word, with what it means and how you'd use it in a sentence. A word that gets added gets one spelling in the lexicon, like every other term.
Can I contribute recordings of my voice?
Yes, and this helps more than anything else. Recordings from more speakers teach it more voices and accents.
Get in touch about recording some sentences.
Who makes it?
Daniel Rosehill, in Jerusalem, as a personal project. It is not made or endorsed by NVIDIA. More on the about page.
Something not covered here? Ask.