OpenSLR 25: ASR data for Amharic, Swahili and Wolof

[cdleong]

Adding a Dataset

• Name: SLR25
• Description: Subset 25 from OpenSLR. Other subsets have been added to https://huggingface.co/datasets/openslr, 25 covers Amharic, Swahili and Wolof data
• Paper: https://www.openslr.org/25/ has citations for each of the three subsubsets.
• Data: Currently the three links to the .tar.bz2 files can be found a thttps://www.openslr.org/25/
• Motivation: Increase ASR data for underrepresented African languages. Also, other subsets of OpenSLR speech recognition have been uploaded, so this would be easy.

https://github.com/huggingface/datasets/blob/master/datasets/openslr/openslr.py already has been created for various other OpenSLR subsets, this should be relatively straightforward to do.

[cdleong]

Whoever handles this just needs to:

• [ ] fork the HuggingFace Datasets repo
• [ ] update the existing dataset script to add SLR25. Lots of copypasting from other sections of the script should make that easy. Amharic URL: https://www.openslr.org/resources/25/data_readspeech_am.tar.bz2. Swahili URL: https://www.openslr.org/resources/25/data_broadcastnews_sw.tar.bz2, Wolof URL: https://www.openslr.org/resources/25/data_readspeech_wo.tar.bz2
• [ ] update the data card to include information about SLR25. There's lots of other examples to draw from.
• [ ] add the appropriate language tags to the data card as well. https://www.w3.org/International/questions/qa-choosing-language-tags, or just use sw, am, and wo for consistency.
• [ ] make a pull request to merge your changes back into HuggingFace's repo

[cdleong]

... also the example in "use in datasets library" should be updated. It currently says But you actually have to specify a subset, e.g.

dataset = load_dataset("openslr", "SLR32")

[cdleong]