whisper-medium-amksim

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.9089
• Wer: 40.3433

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: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 5
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.6349	0.83	5	3.7729	73.3906
3.2338	1.67	10	1.4978	69.0987
1.1335	2.5	15	1.1606	97.4249
0.6838	3.33	20	1.0211	66.0944
0.4383	4.17	25	0.9845	65.2361
0.2514	5.0	30	0.9885	61.3734
0.2053	5.83	35	0.9796	76.3948
0.1353	6.67	40	0.9758	49.3562
0.1142	7.5	45	0.9109	60.9442
0.0889	8.33	50	0.9045	41.2017
0.0854	9.17	55	0.9085	42.4893
0.069	10.0	60	0.9089	40.3433

Framework versions

• Transformers 4.24.0.dev0
• Pytorch 1.13.0
• Datasets 2.6.1
• Tokenizers 0.12.1