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

Blog

Introducing Hebrish

· Daniel Rosehill

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, and every dictated message then needs fixing by hand.

Hebrish is English dictation that writes those Hebrew words correctly, in Latin letters, on your own phone or computer. Today it has a home: hebrish.app.

The mission

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

Hebrish is a not-for-profit project. The apps are free, with no ads and no sign-up. The model and the recordings it is trained on are published under CC-BY-4.0, so anyone can use them, check the results or build on them. Transcription runs on your device; your audio is not uploaded to transcribe it.

What is available

  • The Hebrish app for Android (a voice keyboard), Ubuntu and Windows (tray apps with a hotkey). Download.
  • The model: a fine-tune of NVIDIA Parakeet TDT 0.6B v2. The current version is v5. On held-out test sentences, v3 (the latest scored version) has a word error rate of 5.3% against 10.9% for stock Parakeet, and writes 66% of Hebrew terms correctly against 29%. What those numbers mean.
  • The dataset: thousands of read-aloud recordings of English with Hebrew words in them, with transcripts. Dataset.
  • The lexicon: every Hebrew word and Israeli name the sentences cover, each with one romanized spelling. Browse it, sort it like a spreadsheet, or download it. Lexicon.

One voice, for now

Every training recording so far is one speaker's: mine. I am an Irish-accented English speaker living in Jerusalem, and Hebrish works best for voices like that. If yours is very different, stock Parakeet may suit you better, and the apps switch between the two in one tap. More voices is the biggest thing that would make it better for everyone.

How you can help

Daniel Rosehill, Jerusalem

← All posts · RSS feed