Modelpublic

awaxsama/whisper-small-tunisian-arabic

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes66downloads

Model Card

whisper-small-tunisian-arabic

This model is a fine-tuned version of openai/whisper-small on the linagora/linto-dataset-audio-ar-tn-augmented dataset. It achieves the following results on the evaluation set:

—Loss: 0.0041
—Wer: 0.0019
—Normalized Wer: 0.0019
—Normalized Cer: 0.0012

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: 1e-05
—trainbatchsize: 8
—evalbatchsize: 8
—seed: 42
—gradientaccumulationsteps: 4
—totaltrainbatch_size: 32
—optimizer: Use OptimizerNames.ADAMWTORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizerargs=No additional optimizer arguments
—lrschedulertype: linear
—lrschedulerwarmup_ratio: 0.05
—num_epochs: 3.0
—mixedprecisiontraining: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Normalized Cer	Normalized Wer	Wer
2.652	0.2705	500	0.5770	0.2520	0.4847	0.4879
1.6215	0.5411	1000	0.3265	0.1384	0.2841	0.2847
1.056	0.8116	1500	0.1732	0.0920	0.1790	0.1815
0.4798	1.0817	2000	0.0955	0.0476	0.1032	0.1064
0.2969	1.3522	2500	0.0571	0.0259	0.0586	0.0592
0.187	1.6228	3000	0.0334	0.0117	0.0280	0.0287
0.1178	1.8933	3500	0.0178	0.0053	0.0115	0.0115
0.0494	2.1634	4000	0.0101	0.0029	0.0045	0.0045
0.0381	2.4350	4500	0.0059	0.0019	0.0019	0.0012
0.0296	2.7055	5000	0.0049	0.0025	0.0025	0.0013
0.0246	2.9761	5500	0.0041	0.0019	0.0019	0.0012

Framework versions

—Transformers 4.47.1
—Pytorch 2.5.1+cu124
—Datasets 3.2.0
—Tokenizers 0.21.4