Name: English-Hebrew-Codeswitching-Sentences
Creator: Daniel Rosehill
License: https://creativecommons.org/licenses/by/4.0/

What was recorded, and why

Speech recognition gets Hebrew words wrong when they appear in English speech. To teach a model to get them right, it needs examples: audio of someone saying sentences like "I need to pay the arnona this week", with a transcript that spells the Hebrew word the same way every time.

No such dataset existed, so this one was made. The speaker reads short sentences aloud on a phone. Each sentence is written to contain one or two specific Hebrew terms. The transcript is the sentence as written, with the Hebrew in one romanized spelling.

The dataset itself is all code-switched speech. When the model is trained, about 30% of the training audio is plain English: the same speaker's ordinary English plus LibriSpeech. Fine-tuning on code-switched speech alone tends to damage a model's ordinary English, and the plain English keeps it in shape. Details on the model card.

One speaker. Every recording is Daniel Rosehill: born in Ireland, Irish-accented English, living in Jerusalem. A model trained on this learns one voice. The next step is recordings from more people. You can help.

At a glance

3,018Recordings (rows)

342.4Minutes of audio (305.3 train, 37.1 test)

1,749Distinct Hebrew terms (1,429 train, 320 test)

CC-BY-4.0License

Figures as of 2026-10-19. The dataset grows as new recordings are published; it is the subset of Read-Aloud-Voice-Sentences whose categories include code_switch.

Splits

Split	Rows	Minutes	Distinct target terms
train	2,712	305.3	1,429
test	306	37.1	320

Recorded 2026-10-14 to 2026-10-19.

The test split is term-disjoint

No Hebrew term in the test split appears anywhere in the training split. A model scored on the test split can't have memorised those words from training.

That matters because the useful question is whether a model learns how Hebrew words sound in English speech, not just the list it was shown. Hebrish's results report both: terms seen in training and terms never seen. The gap between them is the honest measure of how far it generalises.

Where the sentences came from

Source	Sentences
Generated by LLMs (mostly Claude Sonnet), each reviewed by a second model from a different family	1,544
Written by hand	1,221
Existing datasets: Tatoeba (CC-BY 2.0 FR), Enron, the author's earlier MIT-licensed datasets	253

Sentence text from other sources keeps its attribution, recorded per row in prompt_source.

How it was recorded

Mostly recorded on a phone (OnePlus Nord 3 5G, internal microphone), with a small number of desktop-mic recordings.
Browser echo cancellation, noise suppression and automatic gain were turned off, so the audio is as the microphone heard it.
16 kHz, mono, 16-bit PCM.
The device, environment and conditions columns record the setup for each row.

Columns

Column	What it holds
`text`	The transcript: English, with Hebrew words romanized
`text_he`	A version with the Hebrew words in Hebrew script (some rows only)
`categories`	Tags for the sentence, including `code_switch`
`target_terms`	The Hebrew terms the sentence was written to contain
`prompt_source`	Where the sentence text came from, for attribution
`prompt_origin`	How the sentence was made
`edited`	Whether the sentence text was edited
`duration_s`	Audio length in seconds
`environment`	Recording environment
`conditions`	Recording conditions
`device`	Recording device
`speaker`	Speaker
`speaker_gender`	Speaker gender
`recorded_at`	When it was recorded

Load it

from datasets import load_dataset

ds = load_dataset("danielrosehill/English-Hebrew-Codeswitching-Sentences")
print(ds)

The recordings