Hungarians/whisper-medium-cv16-hu

Whisper medium Hu

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 16.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0875
• Wer Ortho: 6.6934
• Wer: 5.5500

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: 6.25e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Wer Ortho
0.1877	0.33	1000	0.2104	17.8832	20.5799
0.136	0.67	2000	0.1561	13.4717	16.2140
0.1117	1.0	3000	0.1245	13.4198	10.9487
0.0673	1.34	4000	0.1148	12.0107	9.7836
0.0657	1.67	5000	0.1006	10.3547	8.4702
0.0264	2.01	6000	0.0905	9.0931	7.2250
0.0284	2.34	7000	0.0916	8.7137	7.2221
0.0311	2.68	8000	0.0879	8.0242	6.6914
0.0177	3.01	9000	0.0841	7.6960	6.3860
0.0177	3.35	10000	0.0844	7.2173	6.0125
0.0126	3.68	11000	0.0848	7.2052	5.9739
0.0078	4.02	12000	0.0865	7.1179	6.0629
0.0113	4.35	13000	0.0863	6.9312	5.7990
0.0115	4.69	14000	0.0853	7.0185	5.8968
0.0071	5.02	15000	0.0875	6.6934	5.5500

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0