metadata
license: cc-by-nc-nd-4.0
task_categories:
- automatic-speech-recognition
language:
- fr
size_categories:
- 100K<n<1M
The french subset of the dataset Multilingual TEDx. The data uploaded to HF corresponds to the directory fr-fr. The audio files are automatically resampled to 16 kHz.
Configs:
- single_samples (default): all samples taken separately
- max=30s: combine consecutive samples for a period shorter than 30 seconds
- max=10s: combine consecutive samples for a period shorter than 10 seconds
- max: combine all the samples of a TEDx talk
dependencies (only needed for much faster audio decoding):
- ffmpeg: apt install ffmpeg
- ffmpeg-python: pip install ffmpeg-python
@inproceedings{salesky2021mtedx,
title={Multilingual TEDx Corpus for Speech Recognition and Translation},
author={Elizabeth Salesky and Matthew Wiesner and Jacob Bremerman and Roldano Cattoni and Matteo Negri and Marco Turchi and Douglas W. Oard and Matt Post},
booktitle={Proceedings of Interspeech},
year={2021},
}