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.

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, including the imported ehms_* rows, is his voice. They are intended for fine-tuning nvidia/parakeet-tdt-0.6b-v2 on a single speaker's voice, for use with on-device dictation.

The dataset grows over time. Batches are pushed straight from the capture app on the phone.

Recording device. Every phone take was recorded on a OnePlus Nord 3 5G using its internal microphone, with browser echo cancellation, noise suppression and auto gain turned off. The exception is the ehms_* rows (mode: "import", device: null). These were imported from Hebrish/English-Hebrew-Mixed-Sentences, recorded 2025-11-17 on another device, and their transcripts were updated to this dataset's spelling and casing (the original wording is in original_text). Filter on mode != "import" for phone-only data.

Subsets. This is the full, undifferentiated dataset. Two derived datasets carry its phone takes for one kind of data, unchanged (same ids, columns and split), and are updated on every push:

danielrosehill/English-Hebrew-Codeswitching-Sentences: rows tagged code_switch
danielrosehill/Tech-Sentences: rows tagged tech

Used in

This dataset trains danielrosehill/Parakeet-TDT-0.6B-v2-English-Hebrew-Codeswitching. That model runs on-device in Parlance, the project's dictation app, which offers it alongside stock Parakeet: an Android APK (voice keyboard), an Ubuntu / Debian .deb and a Windows .zip.

Models trained on this dataset, and their results

Parakeet v1.0 (Jachnun) and a streaming Nemotron fine-tune were trained on this dataset at revision 550bfcaa0f84 and scored on 2026-10-19 on its term-disjoint test split (387 code-switched clips, never trained on). Code-switching WER: stock Parakeet TDT 0.6B v2 11.3%, PoC-3 6.0%, v1.0 5.1%, Nemotron 5.8% (offline decoding). Held-out terms written correctly: 25.9%, 45.7%, 47.8% and 44.4%. On the public English-Hebrew-Codeswitching-ASR-Benchmark, v1.0 writes 73.3% of the Hebrew terms correctly, against 32.6% for stock v2.

One speaker's voice: every training and test clip is the same speaker, so these numbers rank models on that speaker only. Full results: hebrish.app/benchmarks; all the datasets: hebrish.app/datasets.

Category	What the sentences contain
`code_switch`	English with everyday Hebrew words written in Latin letters. The spellings follow a single canonical convention.
`names`	Jewish and Israeli first names, surnames and Israeli place names, each with one canonical spelling
`email` / `message`	Sentences in the style of dictated emails, WhatsApp messages and notes to self. This is the main use case: email dictation on Android.
`tech`	Software, DevOps, cloud and AI vocabulary
`homophone`	Sentences whose context settles which spelling of a homophone is meant
`general`	Everyday spoken English
`target_word` / `typed`	Sentences generated on demand around one word, or typed in by the speaker

Wanted data

Where the dataset is meant to grow next, from a voice note by the speaker (2026-10-18; audio, transcript; the full list of ideas is in the model card):

Messier acoustic conditions. Most takes were recorded at home. Takes from the street, cafés and the shuk, and with a baby crying nearby, are wanted, even for sentences with no new vocabulary.
More microphones. All phone takes so far come from one OnePlus Nord 3 5G; desktop-microphone takes are to follow.
Pronunciation variants. Hebrew words with two common pronunciations (Diaspora and Israeli) are recorded in both, with one spelling.
Other speakers, later. Female voices, other English accents common in Israel, and Sephardi and Ashkenazi Hebrew pronunciations. Until then every row is the speaker's own voice.
Better condition tags. Tags set on the phone are often missing; correcting them afterwards, possibly with an audio classifier, is planned.

Layout

The dataset uses the standard audiofolder layout, so load_dataset("danielrosehill/Read-Aloud-Voice-Sentences") works directly.

data/train/audio/<id>.wav    16 kHz mono 16-bit PCM
data/train/metadata.jsonl
data/test/...                held-out set, never trained on

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.

Columns

Column	Meaning
`text`	Ground-truth transcript: punctuated and cased, to match Parakeet v2's output style
`categories`, `target_terms`	What the sentence exercises, and which term the evaluation scores
`prompt_source`	Where the sentence text came from (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
`hebrew_pronunciation`	`israeli`, `ashkenazi`, `mixed`, or null when not tagged (see Pronunciation handling)

Filtering by type

A row can have several categories (code_switch, tech, names, long, email, message, homophone, general), so one kind of data can be used on its own:

from datasets import load_dataset
ds = load_dataset("danielrosehill/Read-Aloud-Voice-Sentences")
code_switch = ds["train"].filter(lambda r: "code_switch" in r["categories"])

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.

Licence and attribution

Audio and transcripts: CC-BY-4.0.
Sentence text whose prompt_source starts with tatoeba:<id> comes from Tatoeba under CC-BY 2.0 FR. The id identifies the source sentence: https://tatoeba.org/en/sentences/show/<id>.
Sentence text whose prompt_source starts with generated: was written for this dataset by the named model. A second model from a different family then reviewed it for correct Hebrew usage and natural phrasing. It is released under CC-BY-4.0 with the rest of the dataset.
Sentence text whose prompt_source starts with enron:<message-id number> is a single sentence taken from the Enron email corpus, which the US Federal Energy Regulatory Commission made public and CMU redistributes. Its licence status is inferred (public record); it has never been formally licensed. Sentences were filtered and reviewed to exclude names, personal content and identifying details.
Sentence text from english-hebrew-mixed-sentences and tech-sentences-for-asr-training comes from the author's earlier datasets, Hebrish/English-Hebrew-Mixed-Sentences and danielrosehill/Tech-Sentences-For-ASR-Training, both MIT.