To create this dataset, a dedicated labeling system (similar to Praat, but web-based) was built. The system lets labelers fix the transcript and align each spoken word to the audio, down to 1ms precision (though annotators typically work at ~10ms granularity).

The audio samples were gathered by randomly sampling from several of ivrit-ai's larger, published open datasets. The original samples contain a transcription and word alignment which were not human reviewed, and were not accurate. This effort is to create an accurate set for both the spoken text and, more importantly, the word-level timing.

Because tagging is ongoing, this dataset is expected to grow in sample count and source diversity — republishing later will include more clips.

Goal of the dataset

The tagging tool this dataset comes from was built to hand-mark Hebrew word boundaries in order to build a gold set — a human-verified reference that says which forced aligner (or aligner configuration) to trust, by comparing automatic alignments against it. This published dataset is that gold set (or a growing slice of it), as opposed to the unlabeled pool that feeds the tagging tool.

Source of samples

Every clip is a short excerpt of a longer source recording (see metadata.recording, the id of that source recording) that was uploaded to the tagging server as part of a larger, not-yet-fully-tagged pool described above. Based on the metadata.source field observed in this export, clips come from (at least) these pipelines:

committee — Israeli Knesset Committee Recordings: ivrit-ai/knesset-committees
plenum — Israeli Knesset Plenum Recordings: ivrit-ai/knesset-plenums
recital — Crowd Recital Project: ivrit-ai/crowd-recital

Each clip also carries a segment_quality score (Machine generated based on Whisper avg logprobs of the segment) and a speakers list (speaker identifiers or, for committee/plenum sources, apparent real speaker names as they appear in the source transcript).

Each clip's audio was extracted based on its labeled words' span, then padded with an extra 200ms before the first word and 1s after the last word, to capture potentially wrong cutoffs. Labelers added words that were spoken in the audio but not originally present in the transcription (see Tagging decisions below). The exported text and word list reflect only the words an annotator ultimately kept in their final labeling — the original sample may have had a different set of words, since words that weren't actually spoken were removed and missing spoken words were added.

Samples from committee are derived from conversations, so overlapping speech from multiple speakers is not rare. We chose to keep those clips and annotate them per the guidelines below, so the benchmark also represents these "hard cases", not just single-speaker clean narration.

Tagging decisions

Annotators worked from the following guidance:

How to match text to speech

Convert numbers/values/dates to spoken words (1970, 2.5, 5%, 14:30, etc.).
Add words which are spoken and not in the text.
Remove words which are not spoken, or only partially spoken (the speaker stopped mid-word, or the word cuts off / starts mid-word in the audio sample).
Use the proper text form of words even if the speaker mispronounces or stutters mid-word — annotators do not write "literal speech representations"; the task is to align proper text to audio.
Do not add words for disfluencies or fillers which are not words themselves (e.g., ahh, mmm)

Multiple speakers

When speakers overlap, annotators time the most dominant speaker's sequence and ignore the background/overlapping words of the other speaker(s).
If a word is clearly spoken by a non-dominant speaker and no dominant speaker is talking at that moment, it is still timed, even if the resulting sentence does not make sense as a whole.
If a speaker is interrupted and yields so another speaker clearly takes over, the interruption words are timed too.

Quality control

Before marking a clip done, the annotator plays the entire clip and watches the word highlights to confirm every word is in sync with the audio.

Weaknesses

Small, partial slice of the source pool. At the time of writing, this dataset contains 72 of the 300 clips (24%) uploaded to the tagging server — ~10m of the pool's ~41.3 minutes. The other 76% are awaiting further labeling rounds.
Source imbalance. Of the 72 published clips, 52 (~72%) come from committee, 13 from recital, and 7 from plenum — the published slice is currently dominated by one source's acoustic/domain characteristics (formal parliamentary speech), which may not generalize to other registers of spoken Hebrew. This may or may not reflect the full pool's proportions, since tagging so far has not necessarily sampled the pool evenly across sources.
Overlapping-speech handling is a modeling choice, not ground truth. Per the tagging guidance above, overlapping non-dominant speech is deliberately ignored rather than transcribed/timed, and a "dominant speaker" call is a subjective judgment made by each annotator — this is a documented convention, not an objective measurement.
Real speaker names. For committee/plenum clips, metadata.speakers can contain the actual names of speakers (e.g. Knesset members) as they appear in the source transcript. Those are public figures and those recordings are in the public domain.
Only 2 annotators so far, one label per clip. As of this writing, the published clips were tagged by 2 distinct annotators, and each clip reflects a single annotator's done submission — not a consensus/adjudicated label. Disagreement between annotators who may have worked the same clip at different times is not captured; there is no independent second-pass QA beyond each annotator's own final self-review (see Quality control above).

Credits and Contribution

This dataset, and the tooling around it were created and maintained by Asael Bar-Ilan (Huggingface, LinkedIn) and the Ivrit.ai team.

License

The dataset is released under the ivrit.ai License, which enables broad research and commercial use.

Full license: https://www.ivrit.ai/en/the-license/
FAQs: https://www.ivrit.ai/en/license-faqs/

Downloads last month: 24

Total file size:

19.5 MB