Dataset Card for NbAiLab/NPSC

The Norwegian Parliamentary Speech Corpus (NPSC) is a speech corpus made by the Norwegian Language Bank at the National Library of Norway in 2019-2021. The NPSC consists of recordings of speech from Stortinget, the Norwegian parliament, and corresponding orthographic transcriptions to Norwegian Bokmål and Norwegian Nynorsk. All transcriptions are done manually by trained linguists or philologists, and the manual transcriptions are subsequently proofread to ensure consistency and accuracy. Entire days of Parliamentary meetings are transcribed in the dataset.

This repository contains a version of the NPSC in the 🤗 Dataset Format. Note that the official release of the dataset, which can be found in the repository of the Norwegian Language Bank, contains more information than the version found here, including word-level metadata, metadata about the speakers, and detailed documentation.

Extra Data

This version of the dataset contains extra data. For both the Norwegian Bokmål and the Norwegian Nynorsk part of the corpus it is added a machine generated text fields that attempts to be orthographic correct. In addition the Norwegian Bokmål text is translated to Nynorsk (and the other way around). The extra fields are named sentence_nob and sentence_nno. In addition any text in brackets, like is removed. In most cases this indicates filler words. Additionally "double words" are deleted. This info is not in the original dataset, and is mainly done to make the text closer to the text used in subtitles.

How to Use

# Loads the 16K Bokmål corpus in streaming mode
from datasets import load_dataset
data = load_dataset("NbAiLab/NPSC_orto", config="16K_mp3_bokmaal", streaming=True)

Dataset Summary

The NPSC dataset contains JSON lines with language training data. The data loader will add audio data to this structure. Here is an example json object:

{
"sentence_id": 49853,
"sentence_order": 0,
"speaker_id": 32,
"meeting_date": "20170110",
"speaker_name": "Olemic Thommessen",
"sentence_text": "Stortingets møte er lovlig satt",
"sentence_language_code": "nb-NO",
"text": "Stortingets møte er lovlig satt",
"sentence_nbo": "Stortingets møte er lovlig satt.",
"sentence_nno": "Stortingets møte er lovleg satt.",
"start_time": 320246, 
"end_time": 323590,
"normsentence_text": "Stortingets møte er lovlig satt",
"transsentence_text": "Stortingets møte er lovleg sett",
"translated": 1,
"audio": {"path": "audio/20170110-095504_320246_323590.wav","array": [.......]}
}

Data Fields

Key	Type	Description
sentence_id:	Integer	Unique identifier of the sentence
sentence_order	Integer	A number indicating the order of the sentences in the meeting
speaker_id	Integer	The ID of the speaker. This can be linked to the original dataset containing thorough demographic and dialectal information about the speaker.
meeting_date	String	The date for the meeting in the format yyyymmdd
speaker_name	String	Name of the speaker. All speakers were members of the Norwegian Parliament or members of the Norwegian Government at the meeting date
sentence_text	String	The sentence text. The transcribed text string of the sentence in non-normalized form. This is the text of the manual transcriptions, without any postprocessing (apart from corrections of known errors). It may contain interrupted words, non-standard words and function words with a pronunciation deviating from the written form. Detailed metadata about the words in the sentence can be found in the word-tokenized version of the corpus in the official release of the dataset.
sentence_language_code	String	The language code of the sentence. The following alternatives exists in the file: ['nb-NO'. 'nn-NO', 'en-US']
text	String	sentence text. This is a copy of "sentence_text". It is included here to make it more convenient to interleave with other datasets.
start_time	Integer	The start time of the sentence in milliseconds. This time is relative to the start of audiofile of the entire meeting, which can be accessed in the official release
end_time	Integer	End time. See comment above.
normsentence_text	String	Normalized sentence text. In this version of the transcription, numbers and dates are written in digits on standardized formats, and common abbreviations are used. These modifications to the original transcriptions are produced automatically using normalization grammars
transsentence_text	String	Translated sentence text. Whenever the original transcription is in Bokmål (nb-NO), this field contains a machine-translated version in Nynorsk (nn-NO), and vice versa
translated	Integer	A flag indicating whether a machine-translated version has been produced or not. Sentences in en-US have not been translated
audio	Array	The dataloader will encode the accociated audio files and provide them as an array containing 'path', 'sound array','sampling_rate'

Initial Data Collection

The procedure for the dataset creation is described in detail in our 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. All recordings are of members of Parliament in a public setting, and can be distributed without any restrains.

Dataset Creators and Curators

The content of the dataset was created by the Norwegian Language Bank (Språkbanken) at the National Library of Norway. Javier de la Rosa, Freddy Wetjen, Per Egil Kummervold, and Andre Kaasen all contributed in making this into a HuggingFace Dataset. Thanks to the HuggingFace team for assistance.

License

The sound and the transcriptions are released under the CC-ZERO-license. The curation of the HuggingFace Dataset is released under CC-BY-SA-3-license.

Citation Information

The following article gives detailed information about the corpus. Please refer to the article and this page if you are using this dataset:


@misc{solberg2022norwegian,
      title={The Norwegian Parliamentary Speech Corpus},
      author={Per Erik Solberg and Pablo Ortiz},
      year={2022},
      eprint={2201.10881},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}