ivrit-ai/knesset-plenums

About

This dataset is derived from raw a/v recordings and human-generated protocols of the Knesset (the Israeli house of representatives) plenums 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, alongside proprietary protocols with timestamps. We extract the audio stream, and clean up timestamp mistakes (such as backward jumps, or out-of-order timestamp artifacts). The protocol text is cleaned, and segmented using the cleaned timestamps. We use Whisper + Stable Whisper (See Below) and word-level alignment 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.

About the Alignment Process

The original audio and protocols contained multiple artifacts that make a reliable alignment difficult.

• Wrong timestamps (lagging, or out-of-order)
• Non-verbatim transcription (this is, by design, a mandate of the Knesset protocol - recorders preserve speaker "spirit" while following proper Hebrew grammar rules)
• Mismatched time base between the recording and the protocol.
• Overlapping speech, noise, and shouting from the audience.

To make things worse, audio recordings are very long (up to 50+ hours), and alignment can easily go out of sync and never converge. We took a pragmatic approach to the alignment process, and used Whisper to align the transcript to the audio. Relying on the logic implemented in Stable-Whisper, we force align the transcript to the audio, but detect divergence and try and recover from it iteratively. This approach assumes "lost segments" and possible corruption - but maintains integrity within segments, and scores each segment so researchers can choose the quality they wish to work with. 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 Plenum Date Ranges: Mar 2009 to Feb 2025
• Total recorded duration: ~8825 hours
• Total Plenum Session: ~1550
• Total Audio+Transcript hours within segments above 0.7 quality score - ~4700 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	45.44%	2475.8
0.9	67.77%	1216.8
0.8	79.87%	659.7
0.7	87.26%	402.8
0.6	91.12%	210.1
0.5	93.69%	139.9
0.4	95.04%	73.8
0.3	96.12%	59.0
0.2	97.16%	56.8
0.1	98.81%	89.7
0.0	100.00%	64.9

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 text contains a mix of topics related to all aspects of life, spanning almost a decade and a half, with samples occurring every few days. This provides researchers with a large, relevant, human-generated text corpus that is grounded in specific dates.

Future Work

• The original recordings contained video streams which we ignored completely - these could be useful for future vision tasks.
• The protocols contain rather noisy but mostly accurate speaker labeling - we ignored this data due to the significant effort involved in cleaning it. We believe speaker detection and diarization tasks could benefit from this additional metadata layer.

Structure

Each top level folder in the set represents a single Plenum session. Each session contains the following files:

• audio.m4a - The audio source. The file is derived from the original A/V recording audio stream without further processing.
• raw.transcript.txt - The protocol text, cleaned up: deduplicate whitespace, preserving newlines.
• raw.transcript.timeindex.csv - Time ranges, into the raw transcript text. Format [start_loc, end_loc, timestamp] - This is derived from the protocol timestamps, after cleaning up the data.
• normalized.transcript.txt - Text derived after normalizing for audio tasks. Cleanup includes: square brackets removal, replacing en-dashes with commas, semi-colons with periods. Non-printable unicode hebrew markers are removed. Whitespace deduplication.
• transcript.json - Transcript, timed using the original (cleaned) protocol timestamps. (Contains the timed text of normalized.transcript.txt). A JSON dump of a WhisperResult
• 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.
• metadata.json - Metadata gathered while normalizing the recording session data and derived from the Recording user profile. (See Below)

Why not VTT/SRT ?

Since recordings can easily go above 24 hours - we could not use the standard subtitle formats (which canot represent >24h media) We also wanted to capture word level timings, and alignment probabilities - Thus the choise of JSON as a flexible container. Converting to VTT/SRT should be straight forward assuming you crop the recording 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. This, you need to download the entire folder as a reposotory of files with something like:

from huggingface_hub import snapshot_download
snapshot_download(repo_id='ivrit-ai/knesset-plenums', 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-plenums', 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 [email protected]:datasets/ivrit-ai/knesset-plenums

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": "plenum",
  "source_entry_id": "100068",
  "plenum_id": "100068",
  "plenum_date": "2021-10-18",
  "duration": 27984.07,
  "quality_score": 0.9009,
  "per_segment_quality_scores": [
    {
      "start": 68.22,
      "end": 68.34,
      "probability": 0.1766
    },
    {
      "start": 70.3,
      "end": 71.76,
      "probability": 0.6063
    },
    ...
  ],
  "segments_count": 9861,
  "words_count": 45658,
  "avg_words_per_segment": 4.6302,
  "avg_segment_duration": 1.8485,
  "avg_words_per_minute": 97.8943,
}

• source_type - Always "knesset"
• source_id - Always "plenum"
• source_entry_id / plenum_id - The unique plenum recording id (not the same as the official plenum id in the Knesset records)
• plenum_date - the date the plenum took place on
• 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 much problems to use. And discard downstream.
• 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 of 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 of the contents of this dataset. Columns within the manifest include:

• source_type
• source_id
• source_entry_id
• plenum_id
• plenum_date
• duration
• quality_score
• min_segment_quality_score
• avg_words_per_segment
• avg_segment_duration
• avg_words_per_minute
• segments_count
• words_count

Auxiliary Plenum Metadata Lookup Table

Since the "plenum_id" in the metadata/manifest files represent an internal "recording id" which is hard to match against the official plenum id. Provided in the knesset_plenums_metadata.csv file is a lookup table of the plenum_id to the official plenum id. This will provide exact start-end dates of the plenum session, the Knesset number, and the running sequential number of the plenum session as recorded by the Knesset.