whisper-medium-ar-no_diacritics

This model is a fine-tuned version of openai/whisper-medium on the audiofolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.1762
• Wer: 7.4970

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
• train_batch_size: 24
• eval_batch_size: 24
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1114	1.01	400	0.1300	9.9764
0.0682	2.02	800	0.1157	8.9138
0.0302	3.03	1200	0.1274	8.2645
0.0151	4.04	1600	0.1277	7.7922
0.0104	5.05	2000	0.1304	7.7922
0.0069	6.06	2400	0.1476	8.4416
0.0033	7.07	2800	0.1307	7.7332
0.0026	8.08	3200	0.1425	8.3235
0.001	9.09	3600	0.1530	8.2054
0.0006	10.1	4000	0.1586	7.9693
0.0008	11.11	4400	0.1601	7.6151
0.001	12.12	4800	0.1647	8.0874
0.001	13.13	5200	0.1650	7.7332
0.0001	14.14	5600	0.1671	7.4380
0.0001	15.15	6000	0.1694	7.2609
0.0001	16.16	6400	0.1726	7.4970
0.0002	17.17	6800	0.1744	7.4380
0.0001	18.18	7200	0.1752	7.4970
0.0	19.19	7600	0.1758	7.4970
0.0	20.2	8000	0.1762	7.4970

Framework versions

• Transformers 4.25.1
• Pytorch 1.12.1
• Datasets 2.8.0
• Tokenizers 0.13.2