Hebrish

Get involved

Help Hebrish learn more

Hebrish gets better with more Hebrew words and more voices. Four ways to help, from two minutes to a whole lecture series.

Ways to help

Which recordings help most

Anything with English and Hebrew mixed in the same speech. The more natural, the better:

  • Religious lectures and shiurim given in English with Hebrew and Aramaic terms.
  • Technical talks: meetups, conference sessions, product demos, team calls you have the rights to.
  • Classes and courses: ulpan, university, professional training.
  • Podcasts, interviews and voice notes from English speakers living in Israel.

Any format is fine: audio or video, any file type, one file or a whole archive.

Even better, with a transcript. If you can, include a timestamped transcript from any automatic tool (Whisper, or whatever you already use). It doesn't need to be correct. Getting the Hebrew words wrong is exactly the problem Hebrish solves: the transcripts are corrected in Hebrish's own training pipeline, and the models are retrained on the corrected text.

To share recordings, get in touch with what you have and roughly how much, and send a signed consent form (below) with it.

Send voice samples in bulk through Hugging Face

If you already record for Hebrish in the app and have given consent, there is nothing more to do: your takes are added through your settings in the app.

If you have a batch of your own recordings to send, the way to do it is a pull request on one dataset, Hebrish/Community-Voice-Contributions. It is the only place that takes contributions; pull requests are switched off on every other Hebrish dataset. We review each pull request and add the recordings to the datasets ourselves.

  1. Record short sentences (3 to 20 seconds each) in English, with the Hebrew words written in Latin letters and numbers in digits. WAV or FLAC, mono, 16 kHz or higher, in a quiet room.
  2. Put them in a folder named after your Hugging Face username: contributions/<username>/audio/, with a metadata.jsonl file next to the audio folder, one line per clip. Follow the example in contributions/danielrosehill/.
  3. Each line needs file_name, id, text (exactly what you read), contributor (your username), is_test (false for real recordings) and consent (true: it is your voice and you release it under CC-BY-4.0).
  4. Open the pull request: on the dataset page choose Files and versions, then Contribute, then Upload files, and tick "Open as a pull request". Or from a terminal: huggingface-cli upload Hebrish/Community-Voice-Contributions contributions/<username> contributions/<username> --repo-type dataset --create-pr

Consent forms

Before recordings can be used, the person or organisation that holds the rights signs a short consent form. Both forms are fillable PDFs with fields for names, the organisation, the material and a signature.

How to use them

  1. Download the form and fill it in, on screen or on paper.
  2. Sign it: type your full name in the signature field, add a digital signature in any PDF reader, or print, sign and scan.
  3. Email it to info@hebrish.app.

What the forms let you choose

  • Training only, or also publishing. By default the material is used only to train and test the models. Tick the open-dataset option to also allow it to be published in Hebrish's public datasets under CC-BY-4.0.
  • Credit or anonymity. Be credited by name, or not at all.
  • Withdrawal. You can withdraw at any time by email. The material is then left out of every new training round and removed from the datasets. Models already trained can't be untrained.

The forms are plain-language templates (version 1.0, 2026-10-28), not legal advice. Questions: info@hebrish.app.