Hebrish/Pronunciation-Variants

Pronunciation Variants

Takes in which a word was deliberately said one particular way, from the main dataset, stratified by variant_category. Two kinds: a term from the project's pronunciation registry said in a recorded variant (pronunciation_variants, a list of {term, variant, category}; first case: Parquet, said par-KET like the English floor word rather than the usual tech reading par-KAY), and a Hebrew word read in a non-Israeli way (hebrew_pronunciation of ashkenazi, anglicized or mispronounced). The transcript is always the canonical spelling (Parquet, Shabbat) whichever way it was said: the point is a model that writes the same word for every common pronunciation of it. These takes stay in training in the main dataset too.

variant_category: cognate (a word shared with another language, read the other language's way), loanword-stress, brand, acronym (letters vs a word), ashkenazi (Hebrew word, Ashkenazi reading) or hebrew (Hebrew word, anglicized or mispronounced).

Website: hebrish.app: get the Hebrish keyboard for Android and the desktop apps, see the results, and find every model, dataset and demo in one place.

This is a subset of danielrosehill/Read-Aloud-Voice-Sentences, the full, undifferentiated dataset. It contains every phone recording there with a non-empty pronunciation_variants, or hebrew_pronunciation of ashkenazi, anglicized or mispronounced; each row adds variant_category. Rows are copied unchanged, with the same ids, columns and train/test split, by scripts/sync_subsets.py in the project repository. Do not edit this dataset directly: fix the row in the main dataset and re-sync. The subset grows as new takes are published.

Speaker: Daniel Rosehill, the dataset's author, who was born in Ireland and speaks English with an Irish accent, and who lives in Jerusalem. Every recording is his voice. Deliberate variation: he sometimes changes his speaking speed on purpose (fast and rushed, or slow) and sometimes says a word the way people commonly mispronounce it, to make the models robust to real speech. The transcript is always the sentence as written, so a take whose speed or pronunciation sounds off against its card is usually intentional, not a recording error.

All audio was recorded on a OnePlus Nord 3 5G phone using its internal microphone, with browser echo cancellation, noise suppression and auto gain turned off.

Pronunciation handling

One of the less obvious facets of this kind of code-switching is that the words switched into English do not come from one language with one pronunciation. Some are Modern Hebrew, some are Yiddish, and even within Hebrew there is more than one pronunciation tradition. Modern Israeli Hebrew uses a Sephardi-based pronunciation (Shabbat, Sukkot, brit); many English-speaking Jews use the Ashkenazi pronunciation of the same words, especially religious ones (Shabbos, Sukkos, bris), and one speaker can move between the two in a single conversation. On top of that is the speaker's own accent in English. That makes this a harder target than code-switching with a single donor language.

Methodology. Variants are recorded deliberately: the speaker reads some Hebrew words in more than one pronunciation, so the same term appears in the dataset said in different ways. The transcript follows the sentence's spelling: "Shabbat" read the Ashkenazi way is still written "Shabbat", and only a sentence that spells the Ashkenazi form ("Shabbos") is transcribed that way.

Tagging. For the best results the variants should be tagged, so models can be scored, and the data balanced, per pronunciation. Since 2026-10-19 each take has a hebrew_pronunciation column:

Value	Meaning
israeli	Modern Israeli (Sephardi-based) pronunciation
ashkenazi	Ashkenazi pronunciation
mixed	Both within the one take
null	Not tagged

Rows recorded before then are null, except those whose transcript already spells an Ashkenazi form, which were filled as ashkenazi. Earlier takes that were read in a deliberate variant are still untagged and will be retro-tagged by listening, so a null does not mean Israeli pronunciation.

Pairs. Since 2026-10-20 each row also has variant_groups, the ids from data/pronunciation-variants.json that link the Israeli and Ashkenazi forms of one word (shabbat for Shabbat / Shabbos, tosafot for Tosafot / Tosfos), and variant_spelling: israeli, ashkenazi, mixed, or null when the transcript has no word that is spelled two ways. Sentences written for an Ashkenazi form come pre-tagged ashkenazi in the capture app, so pairs are labelled as they are recorded. Words spelled the same in both traditions (Torah, Pesach, Kiddush) differ only in hebrew_pronunciation.

Layout

data/train-*.parquet   one row per take: `audio` (the FLAC or WAV file, embedded) and every column
meta/train.jsonl       index: the same rows without audio, plus `shard`, `audio_path` and the audio's byte range
meta/schema.json       column types
data/test-*.parquet, meta/test.jsonl   held-out set, never trained on

Audio is 48 kHz mono 24-bit FLAC for rows from 2026-10-22 and 16 kHz mono 16-bit WAV before.

The test split is term-disjoint: a held-out term never appears in train. load_dataset("Hebrish/Pronunciation-Variants") works directly.

Columns

The columns are the same as in the main dataset: text (ground truth, punctuated and cased like Parakeet v2 output), text_he (the same transcript with its Hebrew terms in Hebrew script, for Hebrew-capable models), categories, target_terms, prompt_source, prompt_origin (manual / dataset / generated), edited, duration_s, environment, conditions, device, speaker, speaker_gender, contributor, hebrew_pronunciation, variant_groups, variant_spelling and recorded_at. The main dataset's card describes them.

Licence

CC-BY-4.0, with sentence-text attribution per prompt_source as described in the main dataset's licence section.