Datasets:

Hebrish
/

Model-Checks

license: cc-by-4.0
language:
  - en
  - he
task_categories:
  - automatic-speech-recognition
pretty_name: Model Checks
tags:
  - speech
  - asr
  - fine-tuning
  - single-speaker
  - asr-evaluation
  - human-feedback
  - error-analysis
  - english
  - code-switching
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

Model Checks

Takes checked by a speech recognition model while they were recorded, with the model's transcript, the speaker's verdict on it and, once judged, a second, machine verdict. Three kinds of row, all from the capture app:

Read-sentence checks (mode bank): the speaker read a sentence (from the sentence banks, or a queue of new sentences written from current web topics, prompt_source queue:<site>), the newest fine-tune transcribed it, and the speaker marked it pass or fail. text is the sentence as written (ground truth).
Free-speech corrections (mode correction, category correction): the speaker said anything, the model got something wrong, and text is the speaker's corrected transcript.
Free-speech passes (mode correction, category correction_accepted, check_verdict pass): the speaker accepted the model's transcript, so text equals model_hypothesis. Not training data: an accepted transcript may still contain a mistake that is already being trained on. These rows are in no other subset.

Columns beyond the main dataset's: model_hypothesis (what the model wrote), asr_model (the model and its Hugging Face revision, e.g. finetune:Hebrish/Sabich-Rolling@<sha7>), check_verdict (pass | fail; null on older corrections, which were all errors), check_errors (the words of text the model got wrong), check_error_types (the speaker's failure types: wrong-word, hebrew-to-english, hebrew-spelling, name-brand, number-format, dropped-word, added-word, split-merge), and the machine judgement: ai_verdict (pass | fail), ai_error_types (same keys), ai_notes (one line), ai_model and ai_judged_at (ISO time), all empty until a judging pass fills them. source_url and topic give the web page a queued sentence was written from.

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.

This is a subset of danielrosehill/Read-Aloud-Voice-Sentences, the full, undifferentiated dataset. It contains every phone recording there with a check_verdict (checked while recording) or recorded in Correct mode (mode correction); each row adds the empty ai_* judgement columns. Rows are copied unchanged, with the same ids, columns and train/test split, by scripts/sync_subsets.py in the project repository. Do not edit this dataset directly: fix the row in the main dataset and re-sync. The subset grows as new takes are published.

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 is his voice. Deliberate variation: he sometimes changes his speaking speed on purpose (fast and rushed, or slow) and sometimes says a word the way people commonly mispronounce it, to make the models robust to real speech. The transcript is always the sentence as written, so a take whose speed or pronunciation sounds off against its card is usually intentional, not a recording error.

All audio was recorded on a OnePlus Nord 3 5G phone using its internal microphone, with browser echo cancellation, noise suppression and auto gain turned off.

Layout

data/train-*.parquet   one row per take: `audio` (the FLAC or WAV file, embedded) and every column
meta/train.jsonl       index: the same rows without audio, plus `shard`, `audio_path` and the audio's byte range
meta/schema.json       column types
data/test-*.parquet, meta/test.jsonl   held-out set, never trained on

Audio is 48 kHz mono 24-bit FLAC for rows from 2026-10-22 and 16 kHz mono 16-bit WAV before.

The test split is term-disjoint: a held-out term never appears in train. load_dataset("Hebrish/Model-Checks") works directly.

Columns

The columns are the same as in the main dataset: text (ground truth, punctuated and cased like Parakeet v2 output), text_he (the same transcript with its Hebrew terms in Hebrew script, for Hebrew-capable models), categories, target_terms, prompt_source, prompt_origin (manual / dataset / generated), edited, duration_s, environment, conditions, device, speaker, speaker_gender, contributor, hebrew_pronunciation, variant_groups, variant_spelling and recorded_at. The main dataset's card describes them.

Licence

CC-BY-4.0, with sentence-text attribution per prompt_source as described in the main dataset's licence section.