| --- |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train-* |
| - split: validation |
| path: data/validation-* |
| - split: test |
| path: data/test-* |
| dataset_info: |
| features: |
| - name: audio |
| dtype: audio |
| - name: dialect |
| dtype: string |
| splits: |
| - name: train |
| num_bytes: 137857739325 |
| num_examples: 186366 |
| - name: validation |
| num_bytes: 382155829 |
| num_examples: 2249 |
| - name: test |
| num_bytes: 1569934935 |
| num_examples: 2307 |
| download_size: 63077132529 |
| dataset_size: 139809830089 |
| --- |
| |
|
|
| # ADI-20: Arabic Dialect Identification dataset and models |
|
|
|
|
| <!-- Provide a quick summary of the dataset. --> |
|
|
| Contact : haroun.elleuch@elyadata.com and fethi.bougares@elyadata.com |
|
|
| We present ADI-20, an extension of the previously published ADI-17 Arabic Dialect Identification (ADI) dataset. |
| ADI-20 covers all Arabic-speaking countries’ dialects. It comprises 3,556 hours from 19 Arabic dialects in addition |
| to Modern Standard Arabic (MSA). We used this dataset to train and evaluate various state-of-the-art ADI systems. |
|
|
| The ADI-20 data set is proposed to broaden the dialectal coverage of ADI-17 and include MSA. |
| First, we incorporated MSA to differentiate between regional dialects and the more formal, |
| standardized Arabic. Second, Tunisian and Bahraini dialects were added to the dataset. |
| Overall, ADI-20 contains 68 hours of MSA speech (MSA dev and test are the same as dev and test sets used in ADI5). |
| Regarding the Tunisian dialect, the whole TunSwitch dataset was used, including the code-switched subset, for a |
| combined duration of 142.7 hours. The train and dev splits were also based on TunSwitch splits but were augmented with |
| content scraped from YouTube to reach 2 hours for each. Lastly, Bahraini content is composed exclusively of content |
| sourced from YouTube, with approximately 271 hours for training and 2 hours for validation and testing each. |
|
|
|
|
| If you are willing to use ADI20, you need to download first ADI17. |
|
|
|
|
| The additional files also include the CSV manifests for ADI-20-53h, the sampled training subset used in the paper for controlled-data experiments. |
|
|
|
|
| ## Paper |
| https://arxiv.org/pdf/2511.10070 |
|
|
| ### Dataset Description |
|
|
| <!-- Provide a longer summary of what this dataset is. --> |
|
|
|
|
| - **Curated by:** Haroun Elleuch |
| - **Shared by [optional]:** Fethi Bougares |
| - **Language(s) (NLP)** : Arabic |
|
|
|
|
|
|
| **BibTeX:** |
| ``` |
| @inproceedings{Elleuch_2025, series={interspeech_2025}, |
| title={ADI-20: Arabic Dialect Identification dataset and models}, |
| url={http://dx.doi.org/10.21437/Interspeech.2025-884}, |
| DOI={10.21437/interspeech.2025-884}, |
| booktitle={Interspeech 2025}, |
| publisher={ISCA}, |
| author={Elleuch, Haroun and Mdhaffar, Salima and Estève, Yannick and Bougares, Fethi}, |
| year={2025}, |
| month=Aug, pages={2775–2779}, |
| collection={interspeech_2025} |
| } |
| ``` |
|
|
