Ara-BEST-RQ-300M-6k
Ara-BEST-RQ-300M-6k is a 300M-parameter self-supervised speech representation model for Arabic and Arabic dialects. It is part of the Ara-BEST-RQ family introduced in
Ara-Best-RQ: Multi Dialectal Arabic SSL.
This model was pretrained on the crawled Ara-BEST-RQ dataset: approximately 5.6k hours of Creative Commons Arabic speech collected from publicly available YouTube videos and segmented for self-supervised speech learning.
Model Details
Model Description
Ara-BEST-RQ is a family of Arabic-focused self-supervised learning (SSL) speech models based on the BEST-RQ framework. The models are designed to learn speech representations that transfer well to Arabic speech processing tasks, including automatic speech recognition (ASR) and dialect identification (DID).
This checkpoint corresponds to the 300M variant pretrained on the crawled 6k-hour dataset.
- Model type: Self-supervised speech representation model
- Architecture: Conformer-based BEST-RQ encoder
- Parameters: ~300M
- Training data: Crawled Arabic speech data
- Training hours: ~5,640 hours
- Languages: Arabic, including multiple dialects
- Primary use: Speech representation learning / downstream fine-tuning
Architecture
The 300M Ara-BEST-RQ model uses:
- 24 Conformer encoder layers
- Model dimension: 848
- 8 attention heads
- Feed-forward dimension: 2048
- GELU activations
- Relative position multi-head attention
- Convolutional front-end
- Random projection quantizer with 4096 codebook entries
Training Data
The model was pretrained on the crawled Ara-BEST-RQ dataset.
The released dataset on Hugging Face provides metadata only: YouTube video identifiers and audio segment boundaries. No audio or video files are distributed as part of the dataset.
Evaluation
The paper evaluates Ara-BEST-RQ models on automatic speech recognition and dialect identification tasks. The following results are reported for the Ara-BEST-RQ crawled 300M model.
Automatic Speech Recognition
WER scores on ASR benchmarks:
| Dataset | WER |
|---|
| Common Voice 19.0 Arabic | 18.67 |
| MGB-3 | 30.85 |
| MGB-5 | 54.18 |
| TARIC-SLU | 23.98 |
| Average | 31.92 |
Dialect Identification
Results on ADI-20:
| Split | Accuracy | Weighted F1 |
|---|
| Validation | 97.21 | 97.17 |
| Test | 96.02 | 95.98 |
Usage
This is a self-supervised pretrained model intended to be used as a speech encoder or as an initialization checkpoint for downstream fine-tuning.
For training and fine-tuning recipes, please refer to the official implementation:
1git clone https://github.com/elyadata/AraBEST-RQ
2cd AraBEST-RQ
You can download the checkpoint from Hugging Face using:
1from huggingface_hub import snapshot_download
2
3model_dir = snapshot_download("Elyadata/AraBEST-RQ-300M-6k")
4print(model_dir)
Please refer to the repository configuration and SpeechBrain recipes for the correct model-loading interface.
Fine-tuning with SpeechBrain
To fine-tune this pretrained Ara-BEST-RQ checkpoint in a SpeechBrain recipe, adapt the pretrainer section of your YAML configuration so that it loads both the pretrained model checkpoint and the corresponding normalizer.
Example:
1pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
2 collect_in: !ref <save_folder>
3 loadables:
4 pt_model: !ref <pt_model>
5 normalize: !ref <normalize>
6 paths:
7 pt_model: !ref <pt_model_path>/model.ckpt
8 normalize: !ref <pt_model_path>/normalizer.ckpt
In your downstream recipe, make sure that:
<pt_model> points to the Ara-BEST-RQ pretrained model object used in your training graph.<normalize> points to the normalization module used by the recipe.<pt_model_path> points to the local directory containing model.ckpt and normalizer.ckpt.<save_folder> is the experiment directory where SpeechBrain should collect and manage pretrained components.
This setup allows SpeechBrain to initialize the downstream model from the Ara-BEST-RQ SSL checkpoint before fine-tuning on task-specific data.
Citation
If you use this model, please cite the Ara-BEST-RQ paper:
1@misc{elleuch2026arabestrqmultidialectalarabic,
2 title={Ara-Best-RQ: Multi Dialectal Arabic SSL},
3 author={Haroun Elleuch and Ryan Whetten and Salima Mdhaffar and Yannick Estève and Fethi Bougares},
4 year={2026},
5 eprint={2603.21900},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2603.21900},
9}
