Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark

English-Hebrew Code-Switching ASR Benchmark

Status (2026-10-25): test set v2 is current. It is the set Hebrish uses for its own internal benchmarks, and the one new results are scored on. Still a small, single-speaker benchmark: read the limits below before drawing conclusions. v1 (2026-10-15) stays here unchanged, with the results scored on it.

A frozen, versioned test set for measuring how well speech recognition handles English speech with Hebrew words in it: "pick up some bamba from the makolet", "I have to go to Bituach Leumi tomorrow". The references write each Hebrew word in one canonical romanized spelling, in Latin letters.

• Website (the Parakeet code-switching fine-tune): 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.
• Runner and scorer: github.com/Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark
• Leaderboard: Space: English-Hebrew-Codeswitching-ASR-Leaderboard

This repository is the test set only. It never changes in place: a new test set is a new version in its own folder, and every result records the version and the commit it was scored against.

Version	Folder	Config	Pinned commit	Status
v2 (2026-10-25)	data/v2/	v2 (default)	48ec85a301dc	Current. Domain-balanced code-switching, 179 clips
v1 (2026-10-15)	data/test/	v1	3cb0b2b3b3e4	History. 145 clips; results scored on it stay valid against that commit

from datasets import load_dataset
v2 = load_dataset("Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark", "v2", split="test")

Listen: what Hebrish is and how it's made

A spoken introduction to the project by the person who makes it: why Hebrish exists, why stock speech-to-text gets Hebrew words wrong, how the fine-tunes are trained, the model lines (Sabich, Jachnun, Shakshuka, Meorav), how the training data is recorded, and the plans for the apps.

Listen	hebrish.app/updates/2026-10-22: player, chapters, edited transcript, MP3 download
Author	Daniel Rosehill
Recorded	2026-10-22
Runtime	1:07:03

Chapters (each opens the player at that point): 0:00 Why this voice note · 1:36 Coming to speech to text · 6:36 Why Hebrew words go wrong · 9:23 How the fine-tune works · 12:27 Why Parakeet, and the model plan · 17:44 Disfluencies, shekels, paragraphs · 25:40 Tech vocabulary and Meorav · 29:07 The kitchen sink: Sabich and Jachnun · 33:35 Licensing, accents, noisy recordings · 39:10 One recorder for ASR and TTS · 41:04 Desktop shortcuts and the macro pad · 43:34 Android: Bluetooth and one button · 45:35 Streaming and Shakshuka · 50:28 Long-form transcription · 56:00 Output modes · 63:01 Report a word, and more voices

Test set v2 (current)

179 clips from the held-out test split of danielrosehill/Read-Aloud-Voice-Sentences at commit 2cec2ebc85e9, all phone recordings from 2026-10-14 to 2026-10-25, each tagged with a subset:

subset	Clips	What it is	Used for
clean	129 (14.8 min)	Code-switching sentences, about 10 per Hebrew vocabulary domain, none of them trained on by any Hebrish model	The headline
english_control	25 (2.0 min)	The same speaker's everyday English with no Hebrew: email, general sentences, homophones	English regression check
tech	25 (2.3 min)	English tech and AI dictation with no Hebrew (npm, Proxmox, Claude Code)	Tech vocabulary check, kept out of the headline

Clean clips per domain

The domains are the ones hebrish.app's lexicon uses. Each clip has one domain, assigned from its Hebrew terms (rule below).

domain	Lexicon label	Clips	Minutes	Term occurrences in these clips	of which unseen	Clips available
places	Places & streets	10	1.2	19	12	44
people	People & public figures	10	0.9	12	5	11
politics	Politics & news	9	1.1	15	12	9
religious	Religious life & calendar	10	1.3	14	11	111
food	Food & shopping	10	1.2	18	10	39
health	Health & pharmacy	10	0.9	19	12	39
bureaucracy	Bureaucracy & government	10	1.3	21	12	23
money	Money & work	10	1.0	16	9	23
home	Home, family & school	10	1.2	18	12	37
army	Army & security	10	1.2	15	11	20
transport	Transport & travel	10	1.3	22	10	15
brands	Brands & acronyms	10	1.2	15	11	19
everyday	Everyday talk	10	1.0	17	10	43
Total		129	14.8	221 (187 distinct terms)	137

Clips available is the eligible pool after every filter below; politics had only 9, so it has 9.

How the clean clips were chosen.

• Eligible: the speaker's own phone takes with a transcript, category code_switch, no tech category (code-switching and tech are kept in separate sentences), no special-delivery task (spelling alphabet, singing, whispering, numbers, long-form), not a 2025 desktop-mic import, not a sentence Whisper-Hebrish trained on, and one take per sentence.
• Never trained on: a clip is out if its id or its normalised text is in the training or validation manifest of any Hebrish training run up to 2026-10-25 (22 runs, including Sabich v2.1, v2.0 and v1.0 and Jachnun v1.0). 13 test-split clips were excluded this way. All 13 matched by text (the same sentence re-recorded), none by id.
• Hebrew terms: a clip's hebrew_terms are the terms from its source label that are entries in hebrish.app's lexicon. English tech and product words (JSON, grep, Proxmox, Tuya: v1's labelling issue) are not lexicon entries, so they are not labelled. Hebrew words the lexicon lacks (Maccabi Zahav, Makor Baruch) are left unlabelled too. They stay in the reference text and count towards WER only.
• Domain rule: the most common domain among the clip's terms. people terms count only when the clip has no other kind, because a name is usually just the addressee ("Hi Gal, ..."). A tie goes to the domain with fewer eligible clips, then alphabetically. 63 of the 129 clips have terms from more than one domain. Each term in hebrew_terms also carries its own domain.
• Sampling: in each domain, clips are ordered by sha256("ehcs-v2:" + id). The first pass takes clips that add a term not yet chosen in that domain, then the rest of the order fills up to 10. The English and tech companions are the first 25 of their pools in the same order. A rerun of the builder gives the same set.
• Unseen: a term is unseen when it is absent from the source dataset's train split at the same commit. 137 of the 221 occurrences are unseen.
• References: 2 source transcripts had a variant spelling (heder madregot) where the label is the canonical term. The reference uses the canonical term and keeps the original in source_text. An oracle check (references scored as transcripts) gives 0% WER and 100% on every term metric for every subset and domain.

Columns (v2)

Column	Meaning
file_name	audio/<id>.wav, 16 kHz mono 16-bit. Clips recorded as 48 kHz / 24-bit FLAC (from 2026-10-22) were resampled
id	Same id as in the source dataset
text	Reference transcript: cased and punctuated, canonical romanized Hebrew
subset	clean, english_control or tech
domain	Lexicon domain of a clean clip; null otherwise
hebrew_terms	List of {term, variants, unseen, domain}. term is the lexicon's canonical spelling. variants are accepted alternative romanizations from the Hebrew-Latin-Token-Classifier lexicon. unseen is true when the term is absent from the source train split. Empty for the English subsets
source_text	The original transcript when the reference corrected a spelling, else null
source_audio	Format of the source file (wav 16000 Hz PCM_16, flac 48000 Hz PCM_24)
duration_s, categories, recording, source_dataset	Provenance

data/v2/build-report.json holds every filter count, the per-domain pools and the contamination check.

Test set v1 (history)

145 clips from the test split of Read-Aloud-Voice-Sentences at commit 3bdaa2a50d52, in data/test/, pinned at commit 3cb0b2b3b3e4 of this repository. Its clean headline was 84 clips with 172 term occurrences; the other subsets were contaminated_text (7), contaminated_audio (26), desktop_mic (4) and english_control (24). Its columns are those of v2 without domain and source_audio.

Known issue, fixed in v2: 22 of v1's 139 clean-subset terms were English tech words or product names (JSON, grep, Proxmox, Tuya, ...), 26 of 172 occurrences, which put term recall 2 to 6 points above a Hebrew-only count. v1 was not balanced by domain. Full v1 description: docs/about.md on GitHub.

Speaker and limits

One speaker: Daniel Rosehill, the author, Irish-born, English-speaking, living in Jerusalem. Clips were recorded on a OnePlus Nord 3 5G internal microphone. This is a single-speaker, read-speech benchmark of 129 headline clips, about 10 per domain. It ranks models on this speaker's code-switching; it does not show how a model handles other accents or spontaneous speech. A per-domain score rests on 10 clips and 12 to 22 term occurrences, so treat it as a pointer, not a measurement. Overall, differences of a few term hits are within noise. Domains come from the lexicon by term, not by sentence context, so a few assignments look odd (Shas is filed under brands, machshev too).

Licence

CC-BY-4.0, like the source dataset. v2 is built by scripts/build_testset_v2.py, v1 by scripts/build_testset.py, in the GitHub repository from pinned upstream commits.