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

Dataset

Clip archive

Every recording in the code-switching dataset, one row per clip. Press play to hear it. Nothing is downloaded until you do.

Each clip has a unique ID. It is the id column of its row in English-Hebrew-Codeswitching-Sentences on Hugging Face, so anyone can cite a clip by it. Click an ID to copy it; Link gives an address that opens this page at that clip (/archive#<id>). The archive lists every clip, whether or not it has been used to train a model yet.

The highlighted words are the clip's target terms: the Hebrew words and Israeli names the sentence was written to contain. Clips marked test are in the held-out test split, which is never trained on. More about the splits and columns on the dataset page.

Pronunciation. Hebrew words are deliberately recorded in different pronunciations: Modern Israeli (Sephardi-based) and Ashkenazi, plus Yiddish words. English speakers in Israel use both, and the model should recognise either. Each clip's hebrew_pronunciation column says which: israeli, ashkenazi or mixed. Tagging started on 2026-10-19; older clips are being tagged retroactively, so many still show no tag.

Loading the dataset from Hugging Face…

    Read live from the main branch of the dataset on Hugging Face (data/train/metadata.jsonl and data/test/metadata.jsonl; audio at data/<split>/audio/<id>.wav), so new recordings appear here as they are published. Audio is 16 kHz mono WAV, CC-BY-4.0.