asr-nigerian-pidgin/pidgin-wav2vec2-base-100H

pidgin-wav2vec2-base-960h

This model is a fine-tuned version of facebook/wav2vec2-base on the Nigerian Pidgin dataset. It achieves the following results on the evaluation set:

• Loss: 1.0898
• Wer: 0.3966

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 4
• eval_batch_size: 4
• seed: 3407
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.3949	1.48	500	3.3325	0.9999
2.4656	2.95	1000	1.4727	0.8026
1.1896	4.43	1500	1.0925	0.6252
0.8558	5.91	2000	0.9467	0.5422
0.6427	7.39	2500	0.9856	0.5096
0.5371	8.86	3000	0.9794	0.5093
0.4553	10.34	3500	0.8719	0.4641
0.3921	11.82	4000	0.9344	0.4566
0.3406	13.29	4500	1.0211	0.4550
0.3046	14.77	5000	0.8668	0.4423
0.2651	16.25	5500	1.0384	0.4261
0.244	17.73	6000	1.0437	0.4296
0.2203	19.2	6500	0.9244	0.4228
0.1995	20.68	7000	0.9832	0.4165
0.1838	22.16	7500	1.1455	0.4112
0.1632	23.63	8000	1.1102	0.4102
0.1576	25.11	8500	1.0769	0.4044
0.1388	26.59	9000	1.1008	0.4013
0.1346	28.06	9500	1.0940	0.4000
0.1204	29.54	10000	1.0898	0.3966

Framework versions

• Transformers 4.37.2
• Pytorch 2.0.1+cu117
• Datasets 2.12.0
• Tokenizers 0.15.2

Citation

@misc{rufai2025endtoendtrainingautomaticspeech, title={Towards End-to-End Training of Automatic Speech Recognition for Nigerian Pidgin}, author={Amina Mardiyyah Rufai and Afolabi Abeeb and Esther Oduntan and Tayo Arulogun and Oluwabukola Adegboro and Daniel Ajisafe}, year={2025}, eprint={2010.11123}, archivePrefix={arXiv}, primaryClass={eess.AS}, url={https://arxiv.org/abs/2010.11123}, }