Dataset · speech audio
English–Hebrew Code-Switching Sentences
English sentences with everyday Hebrew words and Israeli names written in Latin letters, read aloud on a phone by one English speaker living in Israel.
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 | 1,248 | 120.1 | 734 | 754 |
test | 121 | 13.9 | 146 | 85 |
Snapshot as of 15 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
Speech recognition trained on English mangles the Hebrew words that English speakers in Israel use every day: Bituach Leumi, Misrad Hapnim, tzaharon, makolet. This dataset is the training and test data for fixing that.
Each row is one take: a short English sentence containing one or more Hebrew words, read aloud and reviewed on the phone before it was published. Hebrew words follow one canonical romanized spelling, and Hebrew proper nouns are capitalised (Yerushalayim, Bituach Leumi).
It is a subset of Read-Aloud Voice Sentences: every take whose categories include code_switch, copied unchanged with the same ids and split. Fixes go into the main dataset and are re-synced.
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.
- Capture app and dataset created Done 14 October 2026
A phone web app in the style of Common Voice: read, record, review, publish in batches.
- Code-switching subset split out Done 14 October 2026
This dataset, plus a browser Space to listen by term.
- v1 training snapshot Done 15 October 2026
Parent dataset pinned at revision b31aeabb5edc: 77 min of code-switching trained the v1 pilot model.
- Transcript normalisation pass Done 15 October 2026
10 rows fixed, including a spelling-rewrite bug that turned "more" into "moreh".
- v2 training snapshot Done 15 October 2026
Parent dataset pinned at revision 3bdaa2a50d52: 120 min of code-switching (1,248 clips) trained the v2 model; test split 121 clips, 14 min.
- PoC-3 training snapshot Done 18 October 2026
Parent dataset pinned at revision b47310621f79: 196 min of code-switching (1,921 clips) trained PoC-3 (published as v3); test split grew to 219 clips.
- PoC-5 training snapshot Done 18 October 2026
Parent dataset pinned at revision fcf5aec16449: 294 min of code-switching (2,829 clips) trained PoC-5 (PoC-4 used 269 min); test split grew to 311 clips, 265 held-out terms.
- v1.0 training snapshot Done 19 October 2026
Parent dataset pinned at revision 550bfcaa0f84: 339.2 min of code-switching trained v1.0 · Jachnun; the held-out split is 387 clips.
- Data collection In progress
Recording continues, weighted toward the priorities below.
- v6 training snapshot Planned
A new pinned revision at 360 min of code-switching train audio, for v6.
What is being recorded next
- The 390 Hebrew terms that stock Parakeet v2 never gets right, such as Misrad Hapnim, Bituach Leumi, rofe mishpacha and Pikud Haoref.
- Terms with only one or two takes so far. Most of the 600-plus terms in train appear only once.
- 301 Israeli-life terms found missing by a gap review, with 868 new sentences written for them.
- Plain monolingual English from the same phone, for the English share of the training mix (in the parent dataset).
Speaker and recording
- One speaker: Daniel Rosehill, born in Ireland, English with an Irish accent, living in Jerusalem.
- Recorded on a OnePlus Nord 3 5G internal microphone, with browser echo cancellation, noise suppression and auto gain turned off. A smaller set of older takes came from a ThinkPad desktop mic.
- 16 kHz mono 16-bit PCM WAV.
Layout and columns
data/<split>/audio/<id>.wav with a metadata.jsonl per split. load_dataset("danielrosehill/English-Hebrew-Codeswitching-Sentences") works directly once access is granted.
| Column | What it holds |
|---|---|
text | Ground truth, punctuated and cased like Parakeet v2 output |
target_terms | The Hebrew words and names the sentence was written to exercise |
categories | Always includes code_switch; may add names, message and others |
prompt_source | Where the sentence text came from, for attribution |
duration_s | Clip length in seconds |
device / environment / conditions | Recording phone or mic and the room it was recorded in |
recorded_at | UTC timestamp of the take |
Models trained on it
Related
Licence: CC-BY-4.0, with sentence-text attribution per prompt_source as described on the parent dataset's card.