VladS159/Whisper_medium_ro_VladS_4500_steps_multi_gpu_28_02_2024

Whisper Medium Ro - Sarbu Vlad - multi gpu - 3

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 16.1 + Romanian speech synthesis dataset. It achieves the following results on the evaluation set:

• Loss: 0.0743
• Wer: 6.5649

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: 10
• eval_batch_size: 10
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• total_train_batch_size: 30
• total_eval_batch_size: 30
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 450
• training_steps: 4500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2982	0.22	250	0.1790	15.9224
0.1365	0.43	500	0.1313	12.7038
0.1381	0.65	750	0.1126	11.4201
0.1207	0.86	1000	0.1037	11.1432
0.0579	1.08	1250	0.0931	9.6404
0.0665	1.3	1500	0.0929	9.4822
0.0572	1.51	1750	0.0875	9.4457
0.0556	1.73	2000	0.0825	8.6122
0.0458	1.94	2250	0.0778	8.2836
0.0243	2.16	2500	0.0786	7.9095
0.0197	2.38	2750	0.0795	7.8578
0.0229	2.59	3000	0.0758	7.4714
0.0175	2.81	3250	0.0755	7.3497
0.0109	3.03	3500	0.0751	7.0759
0.0098	3.24	3750	0.0773	7.1094
0.0081	3.46	4000	0.0748	6.7778
0.0087	3.67	4250	0.0754	6.6774
0.0086	3.89	4500	0.0743	6.5649

Framework versions

• Transformers 4.37.2
• Pytorch 2.2.0
• Datasets 2.17.0
• Tokenizers 0.15.1