Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark

English-Hebrew Code-Switching ASR Benchmark (test set v1)

Status: early-stage, work in progress. Known issue: 22 of the 139 clean-subset hebrew_terms labels are English tech words or product names (JSON, grep, Proxmox, Tuya, ...), 26 of 172 occurrences. v1.1 will relabel them; until then, term recall is 2 to 6 points higher than a Hebrew-only count. Details: methodology, limits.

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, and every result records the commit it was scored against.

Subsets

145 clips from the held-out test split of danielrosehill/Read-Aloud-Voice-Sentences at commit 3bdaa2a50d52, each tagged with a subset:

subset	Clips	What it is	Used for
clean	84 (11.8 min)	Phone recordings (2026-10-14/15) of sentences that appear in no published training set	The leaderboard headline
contaminated_text	7	New phone recordings of sentences that Whisper-Hebrish was trained on (different audio, same words)	Reported only in all_code_switch
contaminated_audio	26	Desktop-mic recordings (2025-11-17) that are in Whisper-Hebrish's training set, the same files	Reported only in all_code_switch
desktop_mic	4	Desktop-mic recordings from the same 2025 round that no model trained on	Reported only in all_code_switch
english_control	24 (3.3 min)	The same speaker's English with no Hebrew: email, tech, names, homophones	English regression check

The clean subset holds 172 Hebrew term occurrences (139 distinct terms). 117 of those occurrences are of terms that appear nowhere in the training split of the source dataset (unseen: true), so they measure generalisation to new Hebrew words rather than memorisation.

Columns

Column	Meaning
file_name	WAV, 16 kHz mono
id	Same id as in the source dataset
text	Reference transcript: cased and punctuated, canonical romanized Hebrew
subset	See above
hebrew_terms	List of {term, variants, unseen}. term is the 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 dataset's train split
source_text	Set on the 4 rows whose reference was corrected: the source transcript had a variant spelling (heder madregot, machane Yehuda) where its own label was the canonical term, so no transcript could have scored a hit. The reference here uses the canonical term; source_text is the original
duration_s, categories, recording, source_dataset	Provenance

Speaker and limits

One speaker: Daniel Rosehill, the author, Irish-born, English-speaking, living in Jerusalem. Phone clips were recorded on a OnePlus Nord 3 5G internal microphone. This is a single-speaker, read-speech benchmark of 84 headline clips. It ranks models on this speaker's code-switching; it does not show how a model handles other accents or spontaneous speech. Differences of a few term hits are within noise.

Licence

CC-BY-4.0, like the source dataset. Built by scripts/build_testset.py in the GitHub repository from pinned upstream commits.