Hebrish

Same recording. Two models.

Regular speech recognitionStock NVIDIA ParakeetGuesses English

I still need to hit Khaivut from the Kupah and bring the Karti Sashrai to the MRI at Hadassah.

HebrishSabichWrites the Hebrew

I still need a hitchaivut from the kupa, and bring the kartis ashrai to the MRI at Hadassah.

hitchaivut התחייבות referral guaranteekupa קופה health fundkartis ashrai כרטיס אשראי credit card

  • 3.9×more never-seen Hebrew words right62.0% vs 16.1%
  • 3.5×more Hebrew words right70.1% vs 19.9%
  • 2.8×lower word error rate4.91% vs 13.72%

Verbatim transcripts from the public benchmark, test set v2, scored 2026-10-30. The numbers: Sabich's best so far (2026-10-27) against stock NVIDIA Parakeet, on 129 samples no model was trained on. The apps have had rolling updates since. All the results.

AI speech recognition models optimised for the needs of English speakers in Israel and users throughout the Jewish world.

A free voice keyboard and speech-to-text app. Speak English with Hebrew words in it and Hebrish writes what you said, in any app.

Free for personal and non-commercial use. Commercial licensing

Runs on your device. No account, no upload.

Who it's for

Olim and English speakers in Israel

Daily life in English, with the Hebrew left in: kupa, arnona, makolet.

The Jewish world

English speakers anywhere who say Hebrew words: simcha, bar mitzva, parsha.

Yeshivas

The vocabulary of learning, spelled properly: Gemara shiur, chavruta, halacha, beit midrash.

Apps

Free. The app downloads the model once (about 660 MB), then works offline.

Android Voice keyboard and floating mic · Android 9+, arm64 Install with F-Droid Download APK Google Play: coming soon F-Droid updates on its own. The APK link is always the newest version.
Windows Tray app · portable .zip x64 Ubuntu Tray app · .deb Ubuntu 26.04, amd64

New to F-Droid, or on a computer right now? Step-by-step and QR code. iPhone and Mac are planned: why not yet? Other ways to run it: download page.

Server

Want to host Hebrish? If you're an inference provider, Daniel would love to hear from you.

Get in touch

Setup steps: Hebrish Server.

Models

The Hebrish model family consists of three core fine-tunes of NVIDIA speech models, optimised on domain vocabulary for Israel and the Jewish world. Sabich provides higher accuracy and better sentence boundary and punctuation recognition. Shakshuka is for streaming applications, including real-time captioning. Falafel is for voice commands.

Other models

The apps offer these three. Everything else is on the models page: Meorav for text-to-speech, Sabich's training on other base models for small devices, wearables and desktop GPUs, early experiments such as Jachnun and a proof of concept on Qwen ASR, and accessory models. The Sabich variants and early experiments stay at their last version. You can download them from Hugging Face (GGUF, ONNX and NeMo builds) and use them in other apps.

Made by a grumpy Irish oleh chadash in Jerusalem

After two years of using Whisper, Daniel got tired of reading about his trips to the Micollette.

Illustrated portrait of Daniel Rosehill

Daniel RosehillVoice technology enthusiast and AI systems developer · Jerusalem

Daniel is an AI systems developer, originally from Ireland, who became a huge fan of voice transcription. He has been dictating for hours a day for about two years, and even the leading speech models kept getting two things wrong: the Hebrew words that English speakers in Israel use every day, and tech vocabulary. So he trained his own.

Daniel's story

Evals and benchmarks

Sabich against its base model, stock NVIDIA Parakeet, on a code-switching benchmark of 129 samples that no model was trained on.

3.9×

Hebrew words it has never seen

Sabich correctly transcribes 62.0% of Hebrew words it never saw in its training data. The base model only does 16.1%.

3.5×

All Hebrew words

Sabich spells 70.1% of Hebrew words right, against 19.9% for the base model.

2.8×

Lower WER

Code-switching word error rate: 4.91% for Sabich, 13.72% for the base model. Lower is better.

Sabich is still being improved. These are its best scores so far on the public test set v2, scored 2026-10-27. A later version scored 59.9%, 67.0% and 4.96% on 2026-10-30, and the apps have had rolling updates since, not scored yet. Every number: benchmarks.

See the benchmarks

The technical challenge

Modern speech recognition models like Whisper and Parakeet achieve better than 90% accuracy on ordinary speech, at a cost that was impossible a few years ago. But 90% accuracy doesn't mean anyone knows what a Micollette (makolet) is. The accuracy is very high, and the other 10% is still a big problem when it lands on the words that matter.

Daniel is passionate about closing that last 10%: the gap between speech recognition that's extremely useful and almost ready, and text that's good to go out the door as an email or a ping. There's a lot of work in that gap, and code-switching is one of the challenges these models take on.

What is "code-switching"?

Moving between two languages in one sentence, as English speakers in Israel do all day: Everything for the balcony fix is at Home Center, cheaper than the makolet.

Why is it hard?

Speech recognition models are typically trained to expect one language. People don't talk that way. Immigrant populations, such as English-speaking immigrants to Israel and Hispanic communities in the US, routinely mix languages, and Jewish diaspora communities mix Hebrew, Yiddish and even Arabic into their speech.

That mix is a challenge for models trained on less diverse language data. Speech-to-text models are predictive: they favour what they saw most in training. A word that isn't in the main training corpus gets replaced with a sound-alike that means something totally different, so makolet comes out as "macaulette". Hebrew words also have no single spelling in English letters, so Hebrish has to learn one consistent spelling for each.

But Whisper gets Shabbat and kosher right

Models are improving, and some Hebrew words are represented in the main corpus often enough to come out reliably. That still leaves the very long tail:

Bureaucracy & government
teudat zehut, osek patur, rishayon
Health & pharmacy
mirsham, tzilum, Optalgin
Money & work
arnona, tlush, sechirut
Food & shopping
hafuch, krembo, chatzilim
Home, family & school
gan, bagrut, ulpan
Army & security
miluim, mamad, tzav
Transport & travel
sherut, tzomet, machsom
Places & streets
Emek Refaim, Beit Hakerem, Ein Kerem
Religious life & calendar
siyum, mikve, blech
People & public figures
Yossi, Noam, Keren
Brands & acronyms
El Al, Keshet, Home Center
Politics & news
hafgana, mandatim, Yom Haatzmaut
Everyday talk
yalla, dati, shemesh

Browse them all in the lexicon, or read the technical side.

How it's cooked

Every Hebrish model starts from the same place: a good base model, a recipe of sentences, and one blended training run. Pick an ingredient to see what's in it.

Corn the Sloth in a kippah and headset, filling a pita with fried aubergine at a kitchen counter, with bowls of egg, salad and pickles.
Step 1

The base model

NVIDIA Parakeet, which already knows English very well. Fine-tuning keeps that and teaches it what it gets wrong.

Step 2

The fine-tuning dataset

Sentences written around the words a model needs to learn, read aloud and recorded by theme.

Domain vocabulary

Israel code-switching
Tech recognition

Non-vocabulary

Everyday Hebrew

The words of daily life in Israel, left in Hebrew inside English sentences.

  • kupa
  • arnona
  • makolet
  • tofes
  • teudat zehut
  • mirsham

See it in the dataset index

Street names

50 streets in the word list so far. Sentences built from Israel's national street list, starting with Jerusalem and Tel Aviv, are planned.

  • Agripas
  • Agron
  • Ahad Ha'am
  • Ahuza
  • Allenby
  • Arlozorov
  • Bar Ilan
  • Ben Maimon
  • Bograshov
  • Cardo
  • Derech Beit Lechem
  • Frishman
  • Golda Meir
  • Gordon
  • Hamasger
  • Hanevi'im
  • Hapalmach
  • Hayarkon
  • Herzl
  • Hess
  • Ibn Gabirol
  • Jabotinsky
  • Kanfei Nesharim
  • Keren Hayesod
  • King David Residence
  • King George
  • Levinsky
  • Lilienblum
  • Malchei Yisrael
  • Montefiore
  • Nachalat Binyamin
  • Nordau
  • Pierre Koenig
  • Pinsker
  • Rothschild
  • Sderot Begin
  • Sderot Ben Gurion
  • Sderot Hanassi
  • Shaul Hamelech
  • Shazar
  • Shenkin
  • Shivtei Yisrael
  • Shlomzion Hamalka
  • Shmuel Hanavi
  • Strauss
  • Tchernichovsky
  • Ussishkin
  • Yehuda Hamaccabi
  • Yigal Alon
  • Yirmiyahu

See it in the dataset index

Towns and places

Cities, neighbourhoods and landmarks, in their usual English-letter spelling.

  • Emek Refaim
  • Beit Hakerem
  • Ein Kerem
  • Be'er Sheva
  • Bnei Brak

See it in the dataset index

Jewish names

Israeli and Jewish first names and surnames, in messages and emails.

  • Naftali
  • Yotam
  • Tzlil
  • Bat-El
  • Buskila
  • Ifergan

See it in the dataset index

Famous rabbis

Commentators and teachers, including the acronyms people use for them.

  • Rashi
  • Rambam
  • Ramban
  • Rav Kook
  • Chofetz Chaim

See it in the dataset index

Historical figures

Names from Israeli history, many of them on street signs too.

  • Herzl
  • Golda Meir
  • Jabotinsky
  • Yigal Alon

See it in the dataset index

Jewish books

Jewish religious vocabulary and the names of the texts. Real Tanach verses are planned as a set of their own.

  • Mishna
  • Shulchan Aruch
  • Mishna Berura
  • Tehillim
  • Haggada

See it in the dataset index

Yiddish

Yiddish words in English, always also code-switching.

  • rugelach
  • a gut yohr
  • a freilichen Purim

See it in the dataset index

Tech vocabulary

AI and software terms, commands and acronyms.

  • CLI
  • SSH
  • API
  • JSON

See it in the dataset index

Brand names

Company, product, AI model and open-source names in their official spelling.

  • Anthropic
  • OpenRouter
  • Qwen
  • Kubernetes
  • Nginx

See it in the dataset index

Linux admin

Linux commands, tools and system administration.

  • Linux commands
  • system administration

See it in the dataset index

Date formatting

Dates and times said aloud, written the way they are typed. Numbers and dates are in training; a dedicated date and time set is planned.

  • the second of March at seven thirty p.m. → March 2 at 7:30PM

See it in the dataset index

Currency

Spoken amounts written as typed.

  • fifty shekels → 50 NIS
  • twenty dollars → $20

See it in the dataset index

Web addresses

Addresses said aloud, written as typed.

  • dot, slash, colon → as typed

See it in the dataset index

Step 3

Dataset blending and training

The groups are blended into one training set, with plain English mixed in so the model keeps what it knew. The latest published blend, from 2026-10-30, ran to 1,296.3 minutes of training audio:

  • Code-switching 584.9 min
  • Priority words 153.9 min
  • Tech vocabulary 214.9 min
  • Other English 223.2 min
  • LibriSpeech replay 119.4 min
Step 4

The models

Sabich for dictation, Shakshuka for streaming, Falafel for voice commands.

Every set behind this recipe is listed in the dataset index. Browse all 3,642 words and names in the lexicon.

Questions

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.

How does speech-to-text work?

It works on your device, on your own phone or computer. That means it also works with no internet connection, which matters most in an emergency.

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.

How does the training work?

Sentences are written around the words a model needs to learn, read aloud and recorded by theme: code-switching, names and places, Yiddish, religious terms, tech vocabulary, numbers and more. Each training run blends those groups, with plain English in the mix so the model keeps what it knew. See the groups and every dataset.

More questions, including phones, the API and using Hebrish in your app: FAQ.

Licensing

Free forever for personal and non-commercial use. Commercial licensing coming soon, with discounts for non-profit and government users.

The road to commercial licences