Hebrish/Expanded-Acronym-Sentences

Expanded Acronym Sentences

Audio dataset, split-out. Rows copied from the master, danielrosehill/Read-Aloud-Voice-Sentences, so one task can be browsed alone. Training reads the master, not this copy. Every audio dataset as one tree, with hours: the audio index. How every dataset fits together: the dataset index.

Chat and business shorthand said as letters by one speaker ("by E O D", "F Y I", "I M O") with the transcript writing it out in full ("by end of day", "For your information", "In my opinion"). spoken_text is the script with the abbreviation; text is the expanded target. 16 abbreviations (EOD, EOW, COB, FYI, BTW, FWIW, ICYMI, IMO, IMHO, AFAIK, IIRC, TBD, TBC, WFH, OOO, LMK), 6 sentences each. None of them appears as letters in the companion Acronym-Sentences set, so the two never contradict. prompt_kind is time, aside, opinion, status or request.

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.

Two tabs. sentences lists every ground-truth sentence with a takes count, so takes = 0 shows what has not been recorded. takes has one row per recording with its audio, joined on sentence_id; a sentence can have several takes. load_dataset("Hebrish/Expanded-Acronym-Sentences", "sentences") reads the text; the default config reads the takes.

Takes come from danielrosehill/Read-Aloud-Voice-Sentences, the full dataset, with the same ids and train/test split, by scripts/sync_subsets.py in the project repository. Do not edit this dataset directly. New takes append and only the sentence counts are rewritten.

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

sentences/train.parquet  one row per sentence: `id`, `text`, `takes`, `prompt_ids`, `served` and the sentence columns
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/Expanded-Acronym-Sentences") works directly.

Columns

A take's sentence_id is its row's id in sentences. A sentence with served: false is no longer offered by the app; it is kept because a take used it. 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.