Hebrish/Community-Voice-Contributions

Hebrish Community Voice Contributions

A place for anyone to contribute recordings of their own voice to the Hebrish speech models: English sentences with everyday Hebrew words written in Latin letters ("the arnona bill", "a tor at the kupat cholim"), read aloud.

The models were trained on one speaker. More voices, accents and microphones make them better for everyone.

This is the supported way to send voice samples in bulk. Pull requests are switched off on every other Hebrish dataset; open yours here, and we review it and integrate the recordings into the project datasets downstream. For now it holds one example row (ours, marked is_test: true).

Recording in the Hebrish capture app already? If you record in the app and have given consent, your takes are added through your settings there. You don't need to do anything here.

How to contribute

1. Record a few sentences (WAV or FLAC, mono, 16 kHz or higher). Short sentences, 3 to 20 seconds each, in a quiet room. Hebrew words are written in Latin letters, in English sentences; numbers in digits.

2. Make a folder named after your Hugging Face username:

   contributions/<your-hf-username>/
     audio/<id>.flac
     metadata.jsonl
   

3. Add one line to metadata.jsonl per clip (see the example in contributions/danielrosehill/). Required fields: file_name, id, text, contributor, is_test, consent.

4. Open a pull request on this dataset: Files and versions → Contribute → Upload files, tick "Open as a pull request". Or from the command line:

   huggingface-cli upload Hebrish/Community-Voice-Contributions contributions/<your-hf-username> contributions/<your-hf-username> --repo-type dataset --create-pr
   

We review each pull request, then merge it and integrate it into the datasets.

Columns

Column	Type	Meaning
file_name	string	Path to the audio, relative to your folder (audio/<id>.flac)
id	string	Unique clip id, e.g. <timestamp>_<random>
text	string	Exactly what you read, punctuated and cased
contributor	string	Your Hugging Face username (same as the folder name)
is_test	bool	true for test or example clips, which are never trained on; false for real contributions
consent	bool	true = you are the speaker and release the clip under CC BY 4.0 for training and publication
speaker_name	string	Optional. How you want to be credited
speaker_gender	string	Optional
accent	string	Optional, e.g. "American English, lives in Tel Aviv"
categories	list	Optional, e.g. ["code_switch"]
target_terms	list	Optional. The Hebrew words in the sentence
duration_s, sample_rate	number	Optional
device, microphone, environment, recorded_at, notes	string	Optional

Rules

• Only your own voice, and only with consent: true.
• No private information in the sentences.
• Licence: CC BY 4.0. Contributors are credited by name.

Questions: access@hebrish.app.