For training and developing your models in the closed track, we provide the following datasets, which are publicly available on Hugging Face: The datasets represent a wide range of Arabic varieties and recording conditions, with over 85K training sentences in total. The datasets consist of dialectal, modern standard, classical, and code-switched Arabic speech and transcriptions. All except the Mixat and ArzEn subset are diacritized.

We removed samples containing fewer than 3 words and eliminated punctuations from all datasets to enhance consistency and quality. The resulted dataset contains 57K train and 1.5K for dev samples.

For the closed track, you may use the full train/dev sets or a subset of them (for example, you may wish to use the undiacritized subsets for semi-supervised training or rely only on the diacritized subsets). For the open track, you can use these resources and/or any other resources for training, as long as they don't overlap with the test sets.