Dataset · speech audio · Hebrish
Read-Aloud Voice Sentences
Every take recorded for the Hebrish speech models: read-aloud English sentences with Hebrew words, tech vocabulary, names, emails and messages, recorded on a phone and published in batches straight from the capture app.
Gated on Hugging Face (manual approval): request access to download the audio and metadata.
Where the dataset is now
| Split | Clips | Audio (min) | Distinct terms | Phone clips |
|---|---|---|---|---|
train | 5,807 | 642.0 | 2,688 | 5,230 |
test | 536 | 62.7 | 471 | 498 |
Snapshot as of 21 October 2026, counted from the dataset's metadata.jsonl files. The dataset grows as new takes are published, so Hugging Face may already have more. The test split is term-disjoint: its held-out Hebrew terms never appear in train.
About the dataset
This is the full, undifferentiated dataset behind the Hebrish speech models. Every take is a short sentence read aloud and reviewed on the phone before it was published; the dataset grows as new batches are pushed from the capture app.
A row can carry several categories: code_switch (English with Hebrew words written in Latin letters, one canonical spelling each), names, email and message (the main use case: dictating emails and messages on Android), tech, homophone, general, plus later tags such as religious, yiddish, brand-names, irish-english, currency and disfluency. Filter on categories to use one kind of data on its own.
Two derived datasets carry its phone takes for one kind of data, unchanged (same ids, columns and split) and re-synced on every push: English–Hebrew Code-Switching Sentences (rows tagged code_switch) and Tech-Sentences (rows tagged tech). Rows with Jewish religious vocabulary are also mirrored to Jewish-Religious-Terminology-Sentences.
The test split is term-disjoint for code_switch and tech: a held-out Hebrew or technical term never appears in train, so the evaluation measures generalisation to unseen vocabulary. The public benchmark test set was cut from this split at revision 3bdaa2a50d52.
Hebrew words switched into English do not have one pronunciation: Modern Israeli Hebrew says Shabbat, many English-speaking Jews say Shabbos. Variants are recorded on purpose, and since 2026-10-19 each take is tagged with hebrew_pronunciation; since 2026-10-20 variant_groups links the Israeli and Ashkenazi spellings of one word.
Checkpoints and roadmap
Each training run is pinned to a dataset revision, so every model checkpoint can be traced to the exact data it saw. Planned stages are the current plan, not promises.
- Dataset created Done 14 October 2026
First takes pushed from the phone capture app; renamed Read-Aloud Voice Sentences the same day, with explicit configs for the dataset viewer.
- Desktop-mic round imported Done 14 October 2026
498 recordings from English-Hebrew-Mixed-Sentences (2025-11-17) imported, transcripts updated to this dataset's spelling.
- Transcript normalisation and cleanup Done 15 October 2026
Canonical Hebrew spellings and proper-noun casing; 38 corrupted takes removed.
- Benchmark test set cut Done 15 October 2026
The public English-Hebrew-Codeswitching-ASR-Benchmark (test set v1, 145 clips) was built from the test split at revision 3bdaa2a50d52, the same revision that trained PoC-2.
- Wanted-data list Done 18 October 2026
Where the dataset should grow next, from a voice note by the speaker: conditions, microphones, pronunciation variants, other speakers.
- Pronunciation and Yiddish tagging Done 19 October 2026
hebrew_pronunciation column added; rows with Yiddish words tagged yiddish.
- v1.0 training snapshot Done 19 October 2026
Pinned at revision 550bfcaa0f84 for Jachnun v1.0 and the Nemotron streaming fine-tune; its test split (387 code-switched clips) scored them.
- Religious tag and pronunciation pairs Done 20 October 2026
religious category (mirrored to Jewish-Religious-Terminology-Sentences), variant_groups and variant_spelling; a second normalisation pass wrote shekel amounts as NIS.
- Data collection In progress
Recording continues, weighted toward the wanted-data list below.
What is being recorded next
- Messier acoustic conditions: takes from the street, cafés and the shuk, and with a baby crying nearby, even for sentences with no new vocabulary.
- More microphones: every phone take so far comes from one OnePlus Nord 3 5G; desktop-microphone takes are to follow.
- Pronunciation variants: Hebrew words with two common pronunciations (Diaspora and Israeli) recorded in both, with one spelling.
- Other speakers, later: female voices, other English accents common in Israel, and Sephardi and Ashkenazi Hebrew pronunciations.
- Better condition tags: tags set on the phone are often missing; correcting them afterwards, possibly with an audio classifier, is planned.
Speaker and recording
- One speaker: Daniel Rosehill, the dataset's author, born in Ireland, English with an Irish accent, living in Jerusalem. Every recording, including the imported rows, is his voice.
- Phone takes were recorded on a OnePlus Nord 3 5G internal microphone, with browser echo cancellation, noise suppression and auto gain turned off.
- 498 older takes (mode "import", the ehms_* rows) were imported on 2026-10-14 from English-Hebrew-Mixed-Sentences, recorded 2025-11-17 on another device; their transcripts were updated to this dataset's spelling, with the original in original_text. Filter on mode != "import" for phone-only data.
- 16 kHz mono 16-bit PCM WAV.
Layout and columns
data/<split>/audio/<id>.wav with a metadata.jsonl per split. load_dataset("danielrosehill/Read-Aloud-Voice-Sentences") works directly once access is granted.
| Column | What it holds |
|---|---|
text | Ground truth, punctuated and cased like Parakeet v2 output |
categories | What the sentence exercises: code_switch, tech, names, email, message, homophone, general, religious, yiddish and others; a row can have several |
target_terms | The terms (Hebrew or technical) the sentence was written to exercise, and which the evaluation scores |
prompt_source | Where the sentence text came from: tatoeba:<id>, generated:<model>, enron:<id> or one of the author's earlier datasets (see licence) |
edited | true when the speaker corrected the prompt before or after reading it |
duration_s / environment / device / recorded_at | Recording metadata |
conditions | Recording conditions noted on the phone (background-noise, recorded-outdoors, in-car, ...); missing on the earliest rows |
mode | How the take was made; "import" marks the 2025 desktop-mic rows |
hebrew_pronunciation | israeli, ashkenazi, mixed, or null when not tagged (since 2026-10-19; null does not mean Israeli) |
variant_groups / variant_spelling | Ids of Israeli/Ashkenazi spelling pairs that occur in text, and which tradition the text spells (since 2026-10-20) |
Models trained on it
Related
- Hebrish websiteThe apps, the results, and every model and dataset in one place
Licence: CC-BY-4.0, with sentence-text attribution per prompt_source as described on the parent dataset's card.