Dataset Card for "The_Spoken_Wikipedia_Corpora_Dutch_ASR_Hiidden"

Homepage: Spoken Wikipedia Corpora
Repository: Bitbucket Repository
Paper: Publication at LREC 2016
Leaderboard: Interspeech 2018 Paper
Point of Contact: nats@nats.gitlab.io

Dataset Summary

The Spoken Wikipedia Corpora (SWC) is a collection of aligned spoken Wikipedia articles including articles in Dutch. It includes approximately 210 hours of audio, transcriptions, and metadata. The corpus is licensed under CC BY-NC-SA 4.0 and includes valuable resources for research in speech recognition, prosody, and language studies.

Supported Tasks and Leaderboards

The dataset can be used for:

• Automatic Speech Recognition (ASR)

Languages

• Dutch

Dataset Structure

Data Instances

Each instance in the dataset represents a spoken Wikipedia article segment. It includes the audio file and its transcription.

Example Instance

{
  "file_name": "data/audio_1_0.mp3",
  "transcription": "ziek is een buurtschap in de nederlandse gemeente montferland"
}

Data Fields

• file_name: Path to the audio file
• transcription: The corresponding transcript of the audio

Data Splits

The dataset is not explicitly split into training, validation, and test sets. Users can create their own splits based on their specific needs.

Dataset Creation

Curation Rationale

The SWC aims to provide a rich resource for the development and evaluation of ASR systems and other speech-related research. It leverages the diverse and volunteer-driven content of Wikipedia to ensure a wide variety of topics and speaking styles.

Source Data

Initial Data Collection and Normalization

The data was collected from the Spoken Wikipedia project, where volunteers read Wikipedia articles. The audio was then aligned with the text at the sentence level.

Who are the source language producers?

The source language producers are volunteers from the Wikipedia community who contributed by reading articles aloud.

Cleaning, cropping and processing original audio fragments

The transcriptions didn't align exactly with audio fragments (e.g., "Ziek, plaats, uit Wikipedia de Vrije Encyclopedia en helpentwikipedia.org"). In some fragments, another disclaimer is added to the end of the fragment: "Deze opname en alle tekst vallen onder de Creative Commons gelijkdelen licentie". Both are not added to the original transcriptions. To fix this, the following processing steps were taken:

• Used Whisper v3-large to extract sentences with a timestamp.
• Matched these with the original (normalized) transcripts to only keep original ones based on a similarity score. The matching has been done automatically.
• Cropped the audio files and added the corresponding transcriptions.

Additional Processing

The additional processing steps to clean and align the data were developed and executed by @JasperNL. The script used for processing is available upon request.

Annotations

Annotation Process

Annotations include sentence-level alignments, which link the audio to the corresponding text. The annotations were created using an alignment pipeline detailed in the project's repository.

Who are the annotators?

Annotations were created by the developers of the SWC project using automated tools and verified manually to ensure accuracy.

Personal and Sensitive Information

The dataset does not contain personal or sensitive information. It comprises publicly available Wikipedia articles read by volunteers.

Considerations for Using the Data

Social Impact of Dataset

The dataset enables the development of accessible technologies for individuals who prefer or require audio content, such as the visually impaired or those with reading difficulties.

Discussion of Biases

The dataset may reflect biases present in Wikipedia content and the demographic characteristics of its volunteer readers, which could affect the diversity of speech patterns and topics.

Other Known Limitations

The dataset may have incomplete transcriptions and occasional errors in alignment. Some introductory and disclaimer texts were not included in the original transcriptions and need manual processing.

Additional Information

Dataset Curators

The dataset was curated by the SWC project team.

Licensing Information

The dataset is licensed under CC BY-NC-SA 4.0.

Citation Information

If you use this data, please cite the following paper:

@inproceedings{schweitzer2016spwc,
  title={The Spoken Wikipedia Corpora},
  author={Schweitzer, Arne and Lewin, Florian and Lensch, Armin and Mandera, Paweł and Matuschek, Michael and Schiel, Florian},
  booktitle={Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016)},
  year={2016},
  organization={European Language Resources Association (ELRA)},
  note={Dataset available at Hugging Face: @JasperNL https://huggingface.co/datasets/Jaspernl/The_Spoken_Wikipedia_Corpora_Dutch_ASR_Hiidden)}
}

Contributions

Thanks to @natsuhh for adding this dataset and @JasperNL for the additional processing and alignment