README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: sansad_metadata.csv
task_categories:
  - automatic-speech-recognition
  - voice-activity-detection
language:
  - hi
  - en
size_categories:
  - 1M<n<10M
license: other

Dataset Overview

This dataset contains recordings & transcripts from Indian Parliament during 18K sessions totalling upto 17K hours of ~3.2K speakers from 1992-2024.

This dataset is inspired from similar works for European parliament in facebook/voxpopuli

Half of the audios contain weakly-labeled timestamped transcripts as well.

Use Cases

1. Speaker Diarization: Most sessions have been timestamp labeled with speakers. Half of the sessions have all segemnts labeled with a single speaker.
2. Robust Speech-to-Text: Since these datasets have a lot of background noise, clapping, shouting - they'd be good for training robust Speech to Text models.
3. VAD (Voice Activity Detection): Gaps between continuous subtitle segments & speaker segments can be used for Voice Activity Detection / No Speech detection in noisy environments.

Transcription

Roughly half of the videos and #hours also have transcriptions attached in mixed hindi & english. Specifically ~8.6K videos totalling to 9K hours, corresponding to unique 2222 speakers have their transcriptions.

Transcriptions contain ~100M tokens of hindi/english, as measure with gpt4o tokenizer

We've identified some issues with transcriptions, that we haven't studied in detail.

1. We found some cases where the speaker spoke some Indian language other than english/hindi, and the transcripts were in English. Note that the transcript do not contain any language other than English & Hindi.
2. Transcripts aren't exactly as it is, and are often missing in parts thus making a dataset similar to the weakly-labeled transcripts dataset used for training Whisper.

Usage

1. Metadata: can be directly loaded as follows:

from datasets import load_dataset

sansad_dataset = load_dataset("kalpalabs/sansad")

2. Transcripts

   • You can download and unzip transcripts from subtitles.tar.gz file.
   • The complete transcript for a session is stored in subtitles/{session_id}.srt

3. Audios

   • You can download and unzip audios from video_m3u8.tar.gz file.
   • M3U8 file for a row can be found at video_m3u8/{session_id}_360p.m3u8.
   • You can download the full audio by passing the m3u8 file to yt-dlp
   • Please comply with rules and ratelimits of https://videolibrary.sansad.in/ for doing so.

Appendix

License

Disclaimer

This dataset is a compilation of audio and transcripts publicly available from the official Parliament of India website. The creator of this dataset makes no claim of ownership over the original content and has created this dataset for non-commercial research and informational purposes only.

Users of this dataset are URGED to be aware of and respect the original terms of use.

The data has been collected from the official Parliament of India website(s):https://videolibrary.sansad.in/

The original content is subject to the following terms as stated on the Parliament of India website:

"© 2024 Parliament of India. All rights reserved. The content of this website is sole property of Parliament of India and not to be used for commercial or other purpose by any individual and/or organization without consent of Parliament of India."

This dataset was created by kalpalabs.ai to facilitate easier access and analysis of publicly available parliamentary proceedings for non-commercial research, educational, and public interest purposes.

No modifications have been made to the substance of the original audio or transcripts beyond formatting for dataset usability.

The creator of this dataset (the compilation itself) makes no claim of copyright over this specific compilation and dedicates any potential compilation rights to the public domain (CC0). However, this does not affect the copyright or terms of use of the underlying content, which remains the sole property of the Parliament of India.

By downloading or using this dataset, you acknowledge these terms and agree that you are solely responsible for ensuring your use of the content complies with the requirements of the Parliament of India and applicable laws.

It is recommended that users seeking to use this data for purposes beyond personal, non-commercial research or information contact the Parliament of India for clarification or consent.

This dataset is provided "as-is" without any warranty of any kind, express or implied. The dataset creator is not responsible for any issues arising from the use or interpretation of this data.