Wolof-HuBERT-CTC

This model is a fine-tuned version of soynade-research/Wolof-HuBERT-Base. It achieves the following results on a challenging evaluation set:

• Loss: 0.4031
• Wer: 0.3565

It outperforms HuBERT models by Meta and Orange.

Usage

import torch
from transformers import pipeline

pipeline = pipeline(
    task="automatic-speech-recognition",
    model="soynade-research/Wolof-HuBERT-CTC",
    dtype=torch.float16,
    device=0
)

pipeline("https://huggingface.co/soynade-research/Wolof-HuBERT-CTC/resolve/main/story.wav")

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6414	1.1804	10000	0.5430	0.5107
0.3998	2.3607	20000	0.4524	0.4453
0.3896	3.5411	30000	0.4002	0.4217
0.3129	4.7214	40000	0.3863	0.3971
0.2628	5.9018	50000	0.3912	0.3798
0.2275	7.0822	60000	0.3817	0.3717
0.2031	8.2625	70000	0.3872	0.3639
0.1619	9.4429	80000	0.4062	0.3592

Framework versions

• Transformers 4.56.0.dev0
• Pytorch 2.8.0+cu128
• Datasets 2.20.0
• Tokenizers 0.21.4

How to Cite

If you use this model, please cite:

@misc{sy2025speechlanguagemodelsunderrepresented,
      title={Speech Language Models for Under-Represented Languages: Insights from Wolof}, 
      author={Yaya Sy and Dioula Doucouré and Christophe Cerisara and Irina Illina},
      year={2025},
      eprint={2509.15362},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.15362}, 
}