| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| task_categories: |
| - automatic-speech-recognition |
| pretty_name: Example Audio Sentences (format reference) |
| size_categories: |
| - n<1K |
| tags: |
| - reference |
| - format-standard |
| - hebrish |
| configs: |
| - config_name: takes |
| default: true |
| data_files: |
| - split: train |
| path: data/train-*.parquet |
| - config_name: sentences |
| data_files: |
| - split: train |
| path: sentences/train.parquet |
| --- |
| |
| # Example Audio Sentences |
|
|
| **This is the Hebrish project's reference for the audio dataset format.** It is the example a new audio dataset copies, and |
| `audio.schema.json` beside it is the checkable form. The three rows are real takes of Daniel Rosehill copied from |
| [Read-Aloud-Voice-Sentences](https://huggingface.co/datasets/danielrosehill/Read-Aloud-Voice-Sentences) (recorded 2026-11-04), |
| chosen because the sentences are neutral. The place block (`location`, `location_name`, `location_note`, `latitude`, |
| `longitude`, `location_accuracy_m`) and `notes` are blank, as they are in any public reference row. |
|
|
| - Pair: [Example-Sentence-Bank](https://huggingface.co/datasets/Hebrish/Example-Sentence-Bank) is the reference for the sentence type (text to read). |
| - Standard, column meanings, naming, checks and the list of live datasets that differ: [docs/dataset-format-standard.md](https://github.com/danielrosehill/ASR-Training/blob/master/docs/dataset-format-standard.md) in the ASR-Training repo. |
| - Checker: `python3 scripts/check_dataset_format.py meta/train.jsonl --type audio` (ASR-Training repo, stdlib only; a `.parquet` also works with pyarrow installed). |
| - Card layout: [docs/dataset-card-standard.md](https://github.com/danielrosehill/ASR-Training/blob/master/docs/dataset-card-standard.md). This reference card is kept short on purpose. |
| - Website: [hebrish.app](https://hebrish.app) |
|
|
| ## Tabs |
|
|
| One repo per topic (standard of 2026-11-06). `sentences` has every ground-truth sentence with a `takes` count, so `takes = 0` lists what is unrecorded. `takes` has one row per recording with audio, joined on `sentence_id`. New takes only append; only `sentences/*.parquet` is rewritten to update a count. |
|
|
| ## Layout |
|
|
| `data/train-<UTC timestamp>-<4 chars>.parquet` with the audio embedded (FLAC, 48 kHz, 24-bit, mono); the split comes from the |
| file name. `meta/train.jsonl` holds the same rows without `audio` (what the checker reads). `audio.schema.json`, this README. |
| Name a real set `<Topic>-Sentences`. |
|
|
| ## Columns |
|
|
| Required in every row: `audio` (parquet), `id`, `text`, `duration_s`, `sample_rate`, `bit_depth`, `audio_format`, `categories`, |
| `target_terms`, `prompt_id`, `prompt_source`, `mode`, `speaker`, `speaker_id`, `speaker_gender`, `contributor`, `recorded_at`, |
| `device`, `microphone`, `environment`, `transcript_status`, `pronunciation`. Everything else is optional and typed in |
| `audio.schema.json`; the rows here carry the full live column set so the optional ones are visible (null where unset). |
| Column meanings are in the standard document. |
|
|
| `pronunciation` is a reading guide, `[]` by default, otherwise objects `{term, say, note, status}` with `say` a stress |
| respelling (`par-KAY`) and `status` one of `unreviewed`, `ai-reviewed`, `human-reviewed`. It is not `hebrew_pronunciation` or |
| `pronunciation_variants`, which record how this take was actually said. |
|
|
| ## Licence |
|
|
| CC-BY-4.0. |
|
|
