whisper-medium-swagen-combined-20hrs-model

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

• Loss: 0.3817
• Wer: 0.2290

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: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.7478	0.1238	200	0.7976	0.4823
1.6879	0.2477	400	0.6310	0.3895
1.7394	0.3715	600	0.5531	0.3318
1.4248	0.4954	800	0.5066	0.2981
1.2897	0.6192	1000	0.4650	0.2689
1.357	0.7430	1200	0.4458	0.2738
1.3443	0.8669	1400	0.4393	0.2709
1.2284	0.9907	1600	0.4151	0.2745
0.6748	1.1146	1800	0.4113	0.2520
0.7242	1.2384	2000	0.4233	0.2670
0.6472	1.3622	2200	0.4206	0.2592
0.6917	1.4861	2400	0.3990	0.2790
0.7172	1.6099	2600	0.3972	0.3005
0.6105	1.7337	2800	0.3926	0.2313
0.7442	1.8576	3000	0.3817	0.2290
0.7074	1.9814	3200	0.3849	0.2263
0.2659	2.1053	3400	0.4002	0.2486
0.3132	2.2291	3600	0.3958	0.2228

Framework versions

• Transformers 4.47.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0