Hebrish/Hebrish-Validation

Hebrish Validation

A validation benchmark for the Hebrish English-Hebrew code-switching speech recognition models (Sabich and the other lines). Each row is one continuous read of a fixed script of about one minute, with the script as the exact reference transcript. Takes are recorded in the Hebrish capture app; the device each one was recorded on is in its device column.

It complements the public benchmark test set v2, Hebrish/English-Hebrew-Codeswitching-ASR-Benchmark (short sentences); it does not replace it. Here the model hears long-form speech the way it is dictated: whole paragraphs, ordinary devices and rooms.

Never used for training. These recordings are excluded from the Hebrish training corpus and from every dataset derived from it. The script sentences are not in any sentence bank used for recording training data.

Scripts

Three categories, written in English with Hebrew words in the Hebrish romanization:

Category	What it tests
codeswitch	everyday English with romanized Hebrew terms (kupat cholim, ishur, tzaharon)
tech	English technical vocabulary (Kubernetes, PostgreSQL, TLS)
mixed	both in one script

Transcripts keep to plain ASCII letters, digits and . , ? ! ' : - / % $.

Versioning

Set v1.0 (2026-11-02). Scripts change only with a new set version. script_revision (the first 12 hex digits of the SHA-256 of the reference text) pins the exact text each recording was read from.

Speakers

So far one speaker, Daniel Rosehill (male, Irish-born, living in Israel). Recordings by other speakers are very welcome: contact the Hebrish project through https://hebrish.app/feedback.

Columns

Column	What it is
audio	the recording, mono FLAC (48 kHz / 24-bit from the capture app; see sample_rate, bit_depth)
text	the reference transcript (the script)
script_id	the script, e.g. cs-01, tech-02, mixed-03
category	codeswitch, tech or mixed
set_version	the script set, e.g. v1.0
script_revision	SHA-256 prefix of text
duration_s, sample_rate, bit_depth	audio format
device	the device and browser the take was recorded on
microphone	the input used, where the phone reports it
speaker	who read it
created_at	when it was recorded (UTC)
take_id (= id)	the recording's id. A script can have several takes; an evaluator should use the latest per script and speaker.

Index without the audio: meta/validation.jsonl (each row's shard and the byte range of its FLAC in it).

from datasets import load_dataset
ds = load_dataset("Hebrish/Hebrish-Validation", split="validation")

Long-form config (long)

The same takes joined into longer files, so a model is also scored on minutes of continuous dictation. One row per group and speaker: codeswitch, tech and mixed (that category's four scripts, about 3 to 4 minutes each) and all (all twelve, about 10 minutes). Each uses the latest take of every script, in script order, joined with 1.0 s of silence; text is the scripts' texts in the same order, separated by a blank line. Columns: audio, text, group, set_version, script_ids, take_ids and script_revisions (the takes used, in order), gap_s, duration_s, sample_rate, bit_depth, speaker, built_at. Rebuilt when takes change.

long = load_dataset("Hebrish/Hebrish-Validation", "long", split="validation")

Licence

CC BY 4.0, as the rest of the Hebrish datasets.