whisper_result

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

• Loss: 0.6586
• Wer Ortho: 52.0665
• Wer: 49.1825

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: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 50
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wer Ortho
0.6179	0.03	1000	0.9762	60.3624	62.0464
0.505	0.07	2000	0.8327	54.8387	57.7117
0.4921	0.1	3000	0.7555	59.6111	63.5585
0.576	0.13	4000	0.7034	57.6226	58.9214
0.4169	0.17	5000	0.6763	44.5426	46.5726
0.3827	0.2	6000	0.6462	44.9403	47.0766
0.3509	0.23	7000	0.6331	46.4870	48.8407
0.4012	0.26	8000	0.6170	46.8847	49.3952
0.3634	0.3	9000	0.6864	47.4294	45.3822
0.3721	0.33	10000	0.6659	49.2944	46.4870
0.3198	0.36	11000	0.6586	52.0665	49.1825

Framework versions

• Transformers 4.28.0
• Pytorch 2.4.1+cu121
• Datasets 3.0.1
• Tokenizers 0.13.3