whisper-medium-ar-original

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: 0.1852
• Wer: 14.1086

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
0.115	1.01	400	0.1204	18.6541
0.0774	2.02	800	0.1074	15.5844
0.0438	3.03	1200	0.1160	16.4699
0.0233	4.04	1600	0.1279	15.1122
0.0131	5.05	2000	0.1350	15.5254
0.0051	6.06	2400	0.1455	14.9941
0.0035	7.07	2800	0.1464	14.1677
0.0032	8.08	3200	0.1545	14.8170
0.0013	9.09	3600	0.1623	13.8725
0.0013	10.1	4000	0.1543	13.4002
0.0006	11.11	4400	0.1653	14.1677
0.0006	12.12	4800	0.1699	13.7544
0.0003	13.13	5200	0.1705	13.4593
0.0001	14.14	5600	0.1733	13.6954
0.0002	15.15	6000	0.1768	13.8725
0.0001	16.16	6400	0.1786	13.7544
0.0	17.17	6800	0.1826	13.9906
0.0	18.18	7200	0.1839	14.0496
0.0	19.19	7600	0.1848	14.0496
0.0	20.2	8000	0.1852	14.1086

Framework versions

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