README.md

metadata

license: cc-by-sa-4.0
dataset_info:
  features:
    - name: title
      dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 18070709909
      num_examples: 121082466
  download_size: 9046108313
  dataset_size: 18070709909
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Wikipedia Utterances

Text segments extracted from English Wikipedia, segmented into utterance-length chunks suitable for text-to-speech synthesis.

Dataset Description

This dataset contains ~41M text utterances derived from the wikimedia/wikipedia dataset (20231101.en snapshot). Each row contains:

Field	Type	Description
title	string	The Wikipedia article title
text	string	A text segment (10-4,880 characters)
duration	float	Estimated speech duration in seconds (at 150 WPM)

Processing

The transformation pipeline:

1. Tokenizes Wikipedia articles into paragraphs and sentences using NLTK
2. Combines consecutive sentences targeting 15-30 second utterances
3. Strips bracketed content (parentheses, braces, square brackets)
4. Filters for valid utterances ending in sentence-final punctuation

See transform_wikipedia.py in this repository for the full implementation.

Usage

from datasets import load_dataset

dataset = load_dataset("jspaulsen/wikipedia-utterances", split="train")

License

This dataset is released under CC BY-SA 4.0, consistent with the original Wikipedia content license.

Citation

If you use this dataset, please cite the original Wikimedia source:

@ONLINE{wikidump,
    author = "Wikimedia Foundation",
    title  = "Wikimedia Downloads",
    url    = "https://dumps.wikimedia.org"
}