metadata

license: cc0-1.0
language:
  - 'no'
pretty_name: Destil Stortinget
size_categories:
  - 10K<n<100K

Dataset Card for `NbAiLab/nb_distil_speech_noconcat_stortinget`

Dataset Summary

NbAiLab/nb_distil_speech_noconcat_stortinget is a curated subset of the Stortinget Speech Corpus (SSC), a large-scale Norwegian parliamentary speech dataset. This subset focuses on non-concatenated speech segments and includes automatic transcriptions generated using OpenAI's Whisper model. It is designed to facilitate the development and evaluation of Automatic Speech Recognition (ASR) systems, particularly in Norwegian.

Dataset Structure

Data Fields

Each entry in the dataset comprises the following fields:

id: Unique identifier for the audio segment.
group_id: Identifier grouping related segments.
source: Origin of the audio, e.g., "stortinget".
audio: Path to the audio file.
audio_duration: Duration of the audio segment in seconds.
previous_text: Text preceding the current segment in the original transcript.
text: Official transcript of the audio segment.
text_en: English translation of the transcript (if available).
text_language: Language code of the transcript (e.g., "no").
whisper_transcript: Transcript generated by the Whisper model.
whisper_wer: Word Error Rate (WER) of the Whisper transcription.
wav2vec_wer: WER of the wav2vec transcription.
verbosity_level: Indicator of the verbosity level of the speech.
file: Filename of the audio segment.
channels: Number of audio channels.
frequency: Sampling frequency of the audio.
language: Language code of the audio (e.g., "no").
task: Task type, e.g., "transcribe".

Data Splits

The dataset is divided into the following splits:

train: Approximately 195,000 segments.
validation: Approximately 1,540 segments.
validation_clean_stortinget_no: 697 segments.
validation_stortinget_no: Approximately 1,540 segments.

Note: The dataset does not include a test split.

Dataset Creation

Source Data

Original Corpus

The dataset is derived from the Stortinget Speech Corpus (SSC), which comprises over 5,000 hours of Norwegian parliamentary speech. The SSC includes:

Segments: 724,783
Total Duration: 5,190 hours
Unique Speakers: 729

Each segment is up to 30 seconds long and is accompanied by transcriptions in Norwegian Bokmål and Nynorsk.

Reference

The creation of the original SSC is detailed in:

Solberg, P. E., Beauguitte, P., Kummervold, P. E., & Wetjen, F. (2023, May). A Large Norwegian Dataset for Weak Supervision ASR. In Proceedings of the Second Workshop on Resources and Representations for Under-Resourced Languages and Domains (RESOURCEFUL-2023) (pp. 48-52). https://aclanthology.org/2023.resourceful-1.7/

Processing

This subset was created by:

Selecting non-concatenated speech segments from the SSC.
Generating automatic transcriptions using NB-Whisper Large.
Including both the official and Whisper-generated transcriptions to facilitate comparative studies and filtering.

Users can leverage the provided WER metrics (whisper_wer and wav2vec_wer) to filter and select high-quality transcriptions for their specific use cases.

Intended Uses

This dataset is intended for:

Training and evaluating ASR and TTS systems in Norwegian.
Research on weakly supervised learning for speech recognition.
Comparative studies between human and machine-generated transcriptions.

Licensing Information

The dataset is released under the Creative Commons Zero (CC0) license, allowing unrestricted use, distribution, and reproduction in any medium.

Citation

If you use this dataset, please cite the following work: