Hebrish
Experimental Hebrish is an early-stage proof of concept, trained on a small code-switching dataset in one speaker's voice. First results are promising. Growing the training data is a long-term effort, and the aim is for every version to get better. Roadmap

Brand

How Hebrish sounds and looks

Voice tips, ready-to-use templates, logos and colours, for anyone writing about Hebrish: the project, volunteers, press.

Who we are, in one line

Improving ASR accuracy for code-switched English-Hebrew for English-speaking immigrants in Israel and users throughout the Jewish world.

Short form, for taglines and social cards: Dictation that gets the Hebrew words right.

Hebrish is a not-for-profit project for the public good: free apps, open model and data. Say so when you introduce it.

Voice

Plain, specific and measured. Write like a careful engineer explaining a useful tool to a neighbour: short sentences, concrete words, numbers where there are any, and an honest note of what is not known yet. No hype words, no emojis.

Plain

WriteHebrish writes Hebrew words in English speech in one romanized spelling, on your own device.
NotHebrish leverages cutting-edge AI to revolutionise multilingual voice workflows.

Specific

WriteIt writes arnona, kupat holim and teudat zehut instead of guessing English words.
NotIt understands all your Hebrew.

Lead with numbers

WriteCode-switching word error rate 5.3% for v3, against 10.9% for stock Parakeet.
NotMassively more accurate than other models.

Measured

Writev4 and v5 are published and not benchmarked yet. This is an early, proof-of-concept model.
NotThe latest version is even better. (Not measured.)

Say who it's for

WriteFor English speakers in Israel, and Jewish people anywhere, who drop Hebrew words into English.
NotFor everyone who speaks Hebrew. (It does not transcribe Hebrew sentences.)

Keep the caveat

WriteTrained on one speaker's voice, an Irish-accented English speaker in Jerusalem; other voices are untested.
NotWorks for any accent.

Not-for-profit

WriteFree and open: the apps cost nothing, the model and data are CC-BY-4.0.
NotGet Hebrish today! Limited time!

Romanized, per the lexicon

WriteShabbat, makolet, Bituach Leumi, spelled as in the lexicon.
NotShabbos / Shabat / Hebrew script as model output.

Word list

UseAvoid
Hebrish (capital H)HEBRISH, hebrish, Hebrish AI (except as the email sender name)
the Hebrish app; Android, Ubuntu, WindowsParlance (old name: only for the repo and file names)
English with Hebrew words in it; code-switched English-HebrewHebrew transcription; bilingual model
romanized spelling, in Latin letterstransliteration engine, Hebrew support
word error rate (WER), term accuracy / term recallaccuracy (on its own, without saying of what)
not-for-profit project, free and opencharity, non-profit organisation (it is not registered as one), startup
fine-tune of NVIDIA Parakeet TDT 0.6B v2NVIDIA's model, made with NVIDIA (not affiliated)
on your device; your audio is not uploaded100% private, military-grade, secure

Hebrew words are written in their canonical romanization, the spelling in the lexicon. Hebrew script appears only as a reference gloss, never as something the model writes. Numbers are quoted as published on the model card, never rounded up.

Templates

Fill in the [brackets]; delete what doesn't apply. Check every number against the model card before sending.

Social post

Do you dictate in English and drop in Hebrew words like [makolet], [arnona] or [teudat zehut]? Hebrish is free voice dictation that writes those words properly instead of guessing English ones. It runs on your own phone or computer (Android, Ubuntu, Windows): https://hebrish.app

New model version

Hebrish [vN] is out.

What changed: [one sentence: more training audio / new vocabulary / fix]. It was trained on [N] minutes of code-switched speech, up from [N] in [vN-1].

Results: [code-switching WER X% (stock Parakeet 10.9%), Hebrew terms seen in training written correctly X%] — or: not benchmarked yet; scores will follow on the model card.

As before, the training audio is one speaker's voice (an Irish-accented English speaker in Jerusalem), so results on other voices are untested.

The apps pick up the new model on their next update check. Model card: https://hebrish.app/model · Changelog: https://hebrish.app/changelog

App release note

Hebrish app [0.x.y] for [Android / Ubuntu / Windows]

- [What the user can now do, in one line]
- [Fix, in one line: what went wrong before, and what happens now]

[Settings and downloaded models carry over.] Download: https://hebrish.app/download · Release notes: https://github.com/Hebrish-ASR/Hebrish-Dictation/releases

Outreach / introduction email

Subject: [English-Hebrew code-switching dictation: open model and data]

Hello [name],

I make Hebrish (https://hebrish.app), a not-for-profit project that fine-tunes NVIDIA Parakeet to write the Hebrew words in English speech correctly: [makolet], not "Micollette".

[Why you are writing to them, in one or two sentences.]

On held-out test sentences, v3 has a code-switching word error rate of 5.3%, against 10.9% for stock Parakeet. It is trained on one speaker's voice so far. The model, the recordings and the benchmark are CC-BY-4.0.

[The ask: one specific, small thing.]

Best,
Daniel Rosehill

Call for voice recordings

Hebrish needs more voices.

Every recording the model has learned from so far is one speaker's. To make it work for [American / British / South African / Israeli-accented] English, it needs people to read sentences aloud: English with a Hebrew word or two in them, about [N] minutes in total.

The recordings will be published as open data (CC-BY-4.0) and used to train the next version.

Interested? https://hebrish.app/contact?topic=samples

Newsletter update

Subject: Hebrish: [the one headline]

[One-paragraph summary of the headline: what is new and why it matters to someone who dictates.]

Also new:
- [Item, one line, with a link]
- [Item, one line, with a link]

Numbers: the dataset now has [N] recordings ([N] minutes). [Model scores if measured; otherwise "vN is not benchmarked yet".]

How to help: suggest a word (https://hebrish.app/contact?topic=word) or send recordings (https://hebrish.app/contact?topic=samples).

Thanks for following along,
Daniel

A serif H in paper colour on a rounded green square. Use the files as they are: don't recolour, stretch, outline or add effects, and leave clear space of at least a quarter of the mark's width around it.

Mark, SVG Mark for dark backgrounds, SVG Mark, PNG 256 px Wordmark, SVG (light backgrounds) Wordmark, PNG 1080 px

Colour

Accent green on warm paper is the brand. Gold is an accent of the accent: one mark, underline or badge per surface, never a fill. Each swatch shows the light and the dark theme.

Accent
Light #0f6b5c
Dark #3fbf9f
Background
Light #fbf8f2
Dark #111615
Ink
Light #1d2422
Dark #e8ece9
Muted
Light #5b6662
Dark #9aa7a2
Gold (sparingly)
Light #c8901c
Dark #e0b14a

Type

  • Fraunces 600 and 700: headings and the logo.
  • Inter 400, 500 and 600: body text.
  • עברית Frank Ruhl Libre 500 and 700: Hebrew in running text (glosses).
  • Noto Sans Hebrew 500: עבריש in the logo lockup only.

All four are free (OFL) on Google Fonts. Questions about using the brand: contact.