ivrit-ai/knesset-committees

About

This dataset is derived from raw a/v recordings and human-generated protocols of the Knesset (the Israeli house of representatives) committee sessions as part of the ivrit.ai project. Consider visiting the preview space for this dataset here

Method

Data dumps from the Knesset contain A/V recordings of committee sessions, alongside human-generated protocols. We extract the audio stream, abd produce weakly time stamped segmentation of the protocol text (we refer to this process as the pre-alignment) We use Whisper + Stable Whisper (See Below) and pre-aligned text - to force align the protocol text to the audio. In the alignment process, we also score each word for quality of alignment. Some recordings/protocols are discarded due to poor alignment or data corruption.

Unlike the Knesset Plenums dataset, an effort was made to preserve speaker labeling from the original committee protocols. Speaker names and their corresponding text ranges are captured in dedicated files.

About the Alignment Process

Knesset committee recordings present several challenges that make reliable audio-text alignment difficult:

• Timestamps in the original recordings are often inaccurate, delayed, or out of order.
• Committee protocols are not verbatim transcriptions — they preserve the speaker's intended meaning while conforming to formal Hebrew grammar, meaning the spoken and written text often diverge.
• The audio and protocol text may have a different effective time base.
• Recordings contain background noise, overlapping speech, and silent gaps of varying length. The alignment pipeline addresses these challenges through three successive stages, each building on the previous.

Stage 1 — Coarse anchoring (pre-alignment): Since the official protocol text is accurate but carries no timing information, a Whisper-based transcription of the audio is first produced (prealign.transcript.json). This automatic transcription is imprecise, but its timestamps serve as a rough guide. A Hebrew-aware text fingerprinting approach then matches passages of the official text against the automatic transcription to establish loose temporal anchors, and the protocol is split into coarsely timed segments.

Stage 2 — Forced alignment: The coarsely timed segments are passed through a dedicated alignment model that pins each word to its precise position in the audio. Each resulting segment receives a quality score derived from per-word alignment confidence. Researchers can use this score to filter the data to their needs: above 0.5 median score per segment yields generally intelligible results; the project's own downstream tasks use a threshold of 0.7.

Stage 3 — Boundary refinement: Segment boundaries produced by forced alignment can fall mid-word or during active speech. A voice-activity detection (VAD) pass is used to locate the nearest natural silence within each gap and snap boundaries there, ensuring every audio segment begins and ends cleanly.

The pipeline is designed to tolerate imperfect inputs — it assumes some segments will be lost or corrupted — while preserving integrity within successfully aligned segments and providing per-segment quality scores so users can calibrate the trade-off between coverage and accuracy. We find that above 0.5 median score per segment (calculated from word-level probabilities), produces intelligible transcripts. But we choose to work with > 0.7 in our downstream tasks. The code implementing this alignment is part of the main Ivrit.ai data processing repo here (See sources/knesset)

Current Dataset Stats

• Processed Session Date Ranges: Nov 2018 to Jan 2026
• Total recorded duration: ~15,466 hours
• Total Committee Sessions: ~13,177
• Total Audio+Transcript hours within segments above 0.7 quality score: ~7,886 hours (See Below)

Breakdown of segment total duration by quality score cut-offs

Segment Quality Score	% down to quality	Duration (h) at quality
1.0	56.90%	4487.2
0.9	75.03%	1429.8
0.8	84.11%	715.6
0.7	89.67%	438.7
0.6	92.50%	223.3
0.5	94.53%	159.7
0.4	95.62%	86.0
0.3	96.60%	77.8
0.2	97.71%	87.6
0.1	99.27%	122.4
0.0	100.00%	57.8

Segment quality distribution, presenting the quality scores, cumulative percentages, and hours up to each quality level.

How can this dataset be useful?

• Audio tasks - Our main goal is to facilitate audio-related tasks (ASR, TTS) — hence the emphasis on aligned transcriptions.
• Text/NLP tasks - Each session's raw protocol text covers a wide range of legislative, policy, and societal topics, spanning over seven years, with sessions occurring frequently. This provides researchers with a large, relevant, human-generated text corpus grounded in specific dates and committee contexts.
• Speaker diarization / attribution - Unlike the plenums dataset, speaker labels are preserved. This makes the dataset suitable for speaker diarization research and speaker attribution tasks.

Future Work

• The original recordings contain video streams which are currently ignored — these could be useful for future vision tasks.
• Speaker names are preserved as they appear in the protocol - and were not canonicalized or linked back to the Knesset metadata for those speakers.

Structure

Each top-level folder in the dataset represents a single committee session. Each session contains the following files:

• audio.m4a — The audio source. Derived from the original A/V recording audio stream without further processing.
• <id>.doc — The raw protocol document as downloaded from the Knesset committee data sources. Note: these files are actually DOCX files with a .doc extension.
• raw.protocol.txt — The protocol text (the actual spoken content), cleaned up: deduplicated whitespace, preserving newlines.
• raw.full.protocol.txt — The full protocol including the preamble (list of speakers, topics, dates, etc.), not just the spoken content.
• speakers.txt — Speaker names extracted from the protocol.
• speakers.segments.txt — Text ranges (into the raw protocol) attributed to each speaker.
• prealign.transcript.json — A weakly-anchored Whisper transcription of the audio. Used to provide loosely timed anchors as a basis for reliable force-alignment of the accurate protocol text.
• transcript.aligned.json — Transcript, force aligned, with word-level timing and "matching probability" scores. A JSON dump of a WhisperResult.
  • The WhisperResult JSON has a segments top-level property which is a superset of the segments structure the Whisper model produces internally.
• transcript.aligned.map.json — A JSON array with one entry per segment in transcript.aligned.json. Each entry maps that segment back to a character range in raw.protocol.txt and to one or more speaker IDs from speakers.txt. Structure:

  [
    { "start_char": 0,    "end_char": 42,   "speaker_ids": [4762] },
    { "start_char": 42,   "end_char": 86,   "speaker_ids": [4762] },
    { "start_char": 1575, "end_char": 1627, "speaker_ids": [10000000, 5726] },
    ...
  ]
  

  start_char / end_char are byte offsets into raw.protocol.txt. speaker_ids references the numeric IDs in speakers.txt. IDs with values ≥ 10000000 are synthetic identifiers assigned to speakers that could not be matched to a named entry in the protocol preamble.
• transcript.refined.json — A post-processing step using VAD to refine segment start/end timestamps. A JSON dump of a WhisperResult.
• transcript.refined.map.json — Same structure as transcript.aligned.map.json, but referencing segments in transcript.refined.json.
• metadata.json — Metadata gathered while normalizing the recording session data. (See Below)

Why not VTT/SRT?

Since recordings can be very long, we could not use standard subtitle formats (which cannot represent >24h media). We also wanted to capture word-level timings and alignment probabilities — thus the choice of JSON as a flexible container. Converting to VTT/SRT should be straightforward assuming the recording is cropped to less than 24h.

How to download this dataset?

This dataset is stored as a filesystem-repository and does not use the Arrow/Parquet based columnar storage standard of the HF Dataset entity. Thus, you need to download the entire folder as a repository of files with something like:

from huggingface_hub import snapshot_download
snapshot_download(repo_id='ivrit-ai/knesset-committees', repo_type='dataset')

If you encounter timeout errors (usually when downloading from outside AWS) you can increase the timeout and reduce the number of workers to stabilize the download process. e.g.:

snapshot_download(repo_id='ivrit-ai/knesset-committees', repo_type='dataset', etag_timeout=60, max_workers=1)

If this still fails, you can use git to clone the repo (requires installing git LFS). See here

In short, you will need to:

• Install git LFS for your environment + initialize it
• Ensure you have an SSH key uploaded to your HF account
• Configure your local SSH (or SSH agent) to serve this key for hf.co
• Issue something like git clone git@hf.co:datasets/ivrit-ai/knesset-committees

Important — you cannot use load_dataset to download this data, and it requires further processing to be used for training/eval tasks.

Metadata JSON structure

Example metadata:

{
  "source_type": "knesset",
  "source_id": "committee",
  "source_entry_id": "2223415",
  "quality_score": 0.9785,
  "per_segment_quality_scores": [
    { "start": 12.4, "end": 15.1, "probability": 0.8821 },
    { "start": 15.1, "end": 18.6, "probability": 0.9134 },
    ...
  ],
  "segments_count": 690,
  "words_count": 5213,
  "avg_words_per_segment": 7.5551,
  "avg_segment_duration": 2.8157,
  "avg_words_per_minute": 28.2606,
  "session_id": "2223415",
  "duration": 11067.712,
  "title": null,
  "session_date": "2024-11-12 13:30:00",
  "language": "he",
  "knesset_num": "25",
  "category": "כספים",
  "committee_type": "ועדה ראשית",
  "committee_name": "ועדת הכספים",
  "main_category_name": null,
  "session_number": "484",
  "start_date": "2024-11-12T13:30",
  "day_of_week": "3",
  "session_type": "רגילה",
  "comments": null,
  "discussed_topics": "הצעת טיוטת צו תעריף המכס..."
}

• source_type — Always "knesset"
• source_id — Always "committee"
• source_entry_id / session_id — The unique session recording ID
• session_date — The date and start time of the committee session (e.g. "2024-11-12 13:30:00")
• start_date — Same date/time in ISO 8601 format (e.g. "2024-11-12T13:30")
• day_of_week — Day of week as a string integer ("0" = Sunday … "6" = Saturday)
• session_number — The committee's internal sequential session number
• session_type — Session type in Hebrew
• committee_name — The full committee name in Hebrew
• committee_type — Committee type in Hebrew
• category — Short category label in Hebrew
• main_category_name — Top-level category name, if available (otherwise null)
• knesset_num — The Knesset number (parliamentary term) as a string
• title — Session title, if available (otherwise null)
• discussed_topics — Free-text description of the topics discussed, as listed in the protocol
• comments — Additional comments from the protocol preamble, if any (otherwise null)
• language — Always "he"
• duration — Recording duration in seconds
• quality_score — Quality of alignment of the text to the audio. We consider less than 0.3 to be data with too many problems to use.
• per_segment_quality_scores — Allows more granular inspection of per-segment scores to perhaps partially discard some of the data. We consider less than 0.7 to be candidates for discarding.
• segments_count / words_count / avg_words_per_segment / avg_segment_duration / avg_words_per_minute — Recording stats

Manifest file

The root of the dataset folder contains a manifest.csv with a subset of the metadata as a list. This allows a quick glance at the contents of this dataset. Columns within the manifest include:

• source_type
• source_id
• source_entry_id
• language
• quality_score
• min_segment_quality_score
• avg_words_per_segment
• avg_segment_duration
• avg_words_per_minute
• segments_count
• words_count
• session_id
• duration
• session_date
• knesset_num
• committee_name
• start_date

Note Sessions in the manifest file with "empty" quality score have been skipped. The raw data is downloaded, the text and speaker data is extracted when possible, but if the audio file contains no detectable script, or is corrupted - no alignment or transcription takes place and the entry is considered "skipped".

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/