metadata

language:
  - pt
license: mit
datasets:
  - jwlang
tags:
  - automatic-speech-recognition
  - speech
  - dataset
viewer: true
dataset_info:
  config_name: pt
  features:
    - name: client_id
      dtype: string
    - name: audio
      dtype: audio
    - name: sentence
      dtype: string
    - name: language
      dtype: string
    - name: split
      dtype: string
  splits:
    - name: train
      num_bytes: 45540263.152
      num_examples: 1004
    - name: test
      num_bytes: 5906119
      num_examples: 126
    - name: val
      num_bytes: 5474884
      num_examples: 125
  download_size: 56777789
  dataset_size: 56921266.152
configs:
  - config_name: pt
    data_files:
      - split: train
        path: pt/train-*
      - split: test
        path: pt/test-*
      - split: val
        path: pt/val-*

JWLang Corpus

Dataset Summary

The JWLang Corpus is a collection of audio and corresponding text data from JW Broadcasting videos available on the jw.org website. It is intended for training and fine-tuning automatic speech recognition (ASR) models, specifically OpenAI Whisper.

Dataset Structure

Number of samples: 10,000
Data format: Audio (WAV) and Text (SRT)
Size: 5 GB

Splits

Split	Number of samples
Train	8,000
Validation	1,000
Test	1,000

Usage

To load and use the dataset:

from datasets import load_dataset

dataset = load_dataset("M2LabOrg/JWLang_Corpus")

Example Data

Example text snippet from the dataset:

{
  "audio": "path/to/audio.wav",
  "text": "Example subtitle text."
}

License

CC BY-SA 4.0

Citation

If you use this dataset, please cite:

@article{jwlang_corpus,
  title={JWLang Corpus for ASR Training},
  author={Michel Mesquita},
  journal={Unpublished},
  year={2024},
}

Contact

For any questions or issues, please contact Michel Mesquita.