NPSC_test / README.md
versae's picture
Update README.md
0c85e09
metadata
annotations_creators:
  - no-annotation
language_creators:
  - found
language:
  - nb
  - 'no'
  - nn
license:
  - cc0-1.0
multilinguality:
  - monolingual
size_categories:
  - 2G<n<1B
source_datasets:
  - original
task_categories:
  - automatic-speech-recognition
  - audio-classification
task_ids:
  - speech-modeling
pretty_name: NPSC
tags:
  - speech-modeling

Dataset Card for NBAiLab/NPSC

Table of Contents

Dataset Description

The Norwegian Parliament Speech Corpus (NPSC) is a corpus for training a Norwegian ASR (Automatic Speech Recognition) models. The corpus is created by Språkbanken at the National Library in Norway.

NPSC is based on sound recording from meeting in the Norwegian Parliament. These talks are orthographically transcribed to either Norwegian Bokmål or Norwegian Nynorsk. In addition to the data actually included in this dataset, there is a significant amount of metadata that is included in the original corpus. Through the speaker id there is additional information about the speaker, like gender, age, and place of birth (ie dialect). Through the proceedings id the corpus can be linked to the official proceedings from the meetings.

The corpus is in total sound recordings from 40 entire days of meetings. This amounts to 140 hours of speech, 65,000 sentences or 1.2 million words.

This corpus is an adaption of the original corpus made for efficiant ASR training. For simplicity and portability, a few of the original datasets features, like the token transcription, is ommitted. You can find the complete dataset at the Resource Catalogue at Språkbanken.

How to Use (This needs to be edited of course)

from datasets import load_dataset
data = load_dataset("nb/NPSC", streaming=True)

Data Fields

Currently there are two versions included in this repo.

Version A

This verison has a short list of the metadata and includes the audio (48k mp3) encoded as a float32 array in the dataset itself.

The current dataloader script is associated with this version.

One line in train.json looks like this:

{
  "sentence_id": 7309,
  "sentence_order": 0,
  "speaker_id": 1,
  "speaker_name": "Marit Nybakk",
  "sentence_text": "Stortingets møte er lovlig satt",
  "sentence_language_code": "nb-NO",
  "text": "Stortingets møte er lovlig satt",
  "start_time": 302650,
  "end_time": 306000,
  "normsentence_text": "Stortingets møte er lovlig satt",
  "transsentence_text": "Stortingets møte er lovleg sett",
  "translated": 1,
  "audio": {
    "path": "audio/20170207-095506_302650_306000.wav",
    "array": [
      24,
      25,
      50,
      (...)
          ],
    "sampling_rate": 48000
  }
}

Version B

This verison does not contain the audio encoded in the dataset. Instead it has the audio files placed in sub-directories. There are currently both samples in clips_48k_wav and clips_16k_mp3. Only the base filename is referred in the dataset. Please not that there are both sentence-based audio clips as well at meeting-based audio clips. The dataset contains referrals to both, the latter referral has start and stop time as well.

One line in the train/metadata.json looks like this:

{
  "meeting_date": "20170207",
  "full_audio_file": "20170207-095506",
  "proceedings_file": "20170207-095506.ref",
  "duration": 4442474,
  "transcriber_id": 1,
  "reviewer_id": 2,
  "data_split": "test",
  "speaker_name": "Marit Nybakk",
  "speaker_id": 1,
  "sentence_id": 7309,
  "sentence_language_code": "nb-NO",
  "sentence_text": "Stortingets møte er lovlig satt",
  "sentence_order": 0,
  "audio_file": "20170207-095506_302650_306000",
  "start_time": 302650,
  "end_time": 306000,
  "normsentence_text": "Stortingets møte er lovlig satt",
  "transsentence_text": "Stortingets møte er lovleg sett",
  "translated": 1
}

Dataset Creation

We are providing a train, dev and test split. These are the same as in the orginal corpus.

Build date: 20012022

Initial Data Collection and Curation

The procedure for the dataset creation is described in detail in the paper.

Statistics

Feature Value
Duration, pauses included 140,3 hours
Duration, pauses not included 125,7 hours
Word count 1,2 million
Sentence count 64.531
Language distribution Nynorsk: 12,8%
Bokmål: 87,2%%
Gender distribution Female: 38,3%
Male: 61.7%

Considerations for Using the Data

This corpus contains speech data and is allowed to be used outside the National Library of Norway for speech recognition technology purposes.

Discussion of Biases

Please refer to our paper.

Dataset Curators

Per Erik Solberg

Freddy Wetjen, Andre Kaasen and Per Egil Kummervold has contributed to porting it to the Hugging Face Dataset format.

Licensing Information

Licensed for use outside the National Library of Norway.

License

CC-ZERO(https://creativecommons.org/publicdomain/zero/1.0/)

Citation Information

We are preparing an article with detailed information about this corpus. Until it is published, please cite out paper discussing the first version of this corpus:

ANDRE: TO BE DONE