Hebrish
Experimental Hebrish is an early-stage project, 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

Datasets

The data, and which model learned from which

Every Hebrish model is trained on recordings from one growing dataset. This page lists the datasets, and shows which revision of the data each model was trained on, so you can see which models are behind.

One speaker's voice. Every recording in these datasets is one person: Daniel Rosehill, an English speaker in Jerusalem with an Irish accent. Help add more voices.

Which model was trained on which data

All current models train on Read-Aloud-Voice-Sentences, at the revision recorded in each training run. The dataset keeps growing, so a model trained on an older revision has not heard the newer recordings.

ModelTrained onData as ofCode-switched speechStatus
Parakeet v1.0 · Jachnun
run r6-en30		Read-Aloud-Voice-Sentences 550bfcaa2026-10-19339.2 minlatest data
Nemotron, checkpoint 2
run r6-nemotron-en30		Read-Aloud-Voice-Sentences 550bfcaa2026-10-19339.2 minlatest data
Whisper Small ACFT
run small-en-en30-acft		Read-Aloud-Voice-Sentences 36845b0d2026-10-17152.5 minbehind not retrained since 2026-10-17
Whisper Hebrish v2 (large-v3-turbo)
run v2-en30		Read-Aloud-Voice-Sentences 36845b0d2026-10-17152.5 minbehind not retrained since 2026-10-17
Whisper Hebrish v1 (large-v3-turbo)English-Hebrew-Mixed-Sentences e7dc91ef2025-11-17516 clipssuperseded by v2

Status checked 2026-10-19: the dataset's latest revision was 550bfcaa (2026-10-19 17:13 UTC). Minutes are the code-switched speech in each training run's mix, as recorded in the run. Parakeet and Nemotron also trained on 95.3 min of the speaker's own English and 48.5 min of LibriSpeech; the Whisper v2 runs on 40.5 and 22.0 min.

What the latest data bought: the two models trained on revision 550bfcaa, scored on its held-out split. Benchmarks · white paper (PDF).

Bar charts comparing stock Parakeet, PoC-3, v1.0 and Nemotron on the held-out split: code-switching WER and term recall.

Recordings

  • Read-Aloud-Voice-Sentences

    The parent dataset: every sentence read aloud for the project, on one phone. Code-switched sentences, plain English, tech vocabulary and Yiddish words. The subsets below are copied from it, row for row.

    CC-BY-4.0 · grows with every recording session

    Trains: every current model (table above).

  • English-Hebrew-Codeswitching-Sentences

    The code-switched subset: English sentences with Hebrew words in them. A held-out list of Hebrew terms is recorded only in its test split.

    CC-BY-4.0 · 3,331 recordings, 375.3 min (revision 7bfcea1d) · Hugging Face · listen to every clip

    Part of: Read-Aloud-Voice-Sentences, so it trains every current model.

  • Yiddish-Sentences

    English sentences with Yiddish words in Latin letters (shul, daven, cholent, tachles), the way English speakers in Israel use them.

    CC-BY-4.0 · subset of Read-Aloud-Voice-Sentences

  • Tech-Sentences

    Dictation-style English dense in software, DevOps and AI vocabulary (GitHub, PostgreSQL, kubectl), in the vendors' own casing.

    CC-BY-4.0 · subset of Read-Aloud-Voice-Sentences

  • English-Hebrew-Mixed-Sentences

    Where it started: 516 clips recorded on a desktop on 2025-11-17, three sentences per Hebrew word. 498 of them were later brought into Read-Aloud-Voice-Sentences with their spellings made canonical, so they are part of what every current model trains on.

    MIT · 2025-11-17

    Trained: Whisper Hebrish v1.

Evaluation

Text and outside data

  • hebrew-latin-code-switching

    Text only, no audio: English sentences with Hebrew words in Latin letters, with token-level spans marking them. For detecting Hebrew words in English text, not for training speech models.

    MIT

  • LibriSpeech

    Public English audiobook speech, mixed into Parakeet and Nemotron training (with cased, punctuated LibriSpeech-PC transcripts) so the models keep their English. Not made by Hebrish.

    CC-BY-4.0 · train-clean-100

The vocabulary these datasets cover is in the lexicon explorer. The models themselves: models and developer notes.