ivrit-ai/crowd-recital

About

This dataset was created by crowd-sourced recording sessions in Hebrew as part of the ivrit.ai Crowd Recital project.

Volunteers read on normal desktop or mobile setting Wikipedia articles while time-stamping every sentence read. Later this data is normalized by aligning the gathered captions with the audio using Stable Whisper (See Below). The recording project is an ongoing effort and new data will be appended to this dataset periodically as it is being generated.

Current Dataset Stats

• Cutoff date: Feb 2024
• Total recorded duration: 50.4 Hours
• Total Sessions: 2398
• Total Words: 325,152
• Total recorders: 85

Structure

Each top level folder in the set represents a single recording session. Each recording session at the following files:

• audio.mka - The audio source. The file is derived from the browser recording session. Common sample rates of 48Khz and 42Khz but in general, they are browser dependent. Audio is encoded using the browser default encoder (Usually Opus for Chrome/Firefox browsers and AAC for Safari)
• transcript.vtt - The captions as they were presented to the recorder times stepping is derived from the recorder clicks on the keyboard alignment may not be accurate and is always on a sentence level.
• transcript.aligned.json - A JSON dump of a WhisperResult structure which is the output of Stable Whisper alignment of the captions with the audio using some underlying Whisper model. -- The WhisperResult json has a "segements" 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.

Metadata JSON structure

Example metadata:

{
  "source_type": "crowd_recital",
  "source_id": "recital",
  "source_entry_id": "Ia4wXanRR9QmFIlkcSYbf",
  "session_id": "Ia4wXanRR9QmFIlkcSYbf",
  "session_duration": 3.2266666666666666,
  "user_id": "7a9c13de-e64b-439c-be41-90230b758204",
  "document_language": "he",
  "document_title": "\u05e9\u05dc\u05de\u05d4 \u05e9\u05e9",
  "document_source_type": "wiki-article",
  "year_of_birth": 1984,
  "biological_sex": "male",
  "quality_score": 0.8715,
  "per_segment_quality_scores": [
    {
      "start": 1.32,
      "end": 2.42,
      "probability": 0.8715
    }
  ],
  "segments_count": 1,
  "words_count": 2,
  "avg_words_per_segment": 2.0,
  "avg_segment_duration": 1.1,
  "avg_words_per_minute": 37.1901
}

• source_type - Always "crowd_recital"
• source_id - Always "recital"
• source_entry_id / session_id - The unique recording session id
• session_duration - recording duration in seconds
• user_id - An Opeque user id - can be used to group recordings by the same user
• document_language - The read text language - Expected to always be "he" for this dataset
• document_title - Hebrew encoded unicode title (Usually derived from the Wiki article title)
• document_source_type - In almost all cases will be "wiki-article". Some sources are uploaded content with safe and clear copyright, those would be "plain-text" or "file-upload"
• year_of_birth - The recorder year of birth if provided
• biological_sex - The recorder biological sex if provided
• 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 0.5 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
• session_id
• session_duration
• user_id
• document_language
• document_title
• document_source_type
• year_of_birth
• biological_sex
• quality_score
• min_segment_quality_score
• avg_words_per_segment
• avg_segment_duration
• avg_words_per_minute
• segments_count
• words_count