VladS159/Whisper_medium_ro_VladS_8000_steps_multi-gpu_11_05_2024

Whisper Medium Ro - Sarbu Vlad - multi gpu --> 3

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 17.0 + Romanian speech synthesis dataset. It achieves the following results on the evaluation set:

• Loss: 0.0777
• Wer: 5.7842

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: 11
• eval_batch_size: 10
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• total_train_batch_size: 33
• 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: 800
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1807	0.47	500	0.1359	13.4050
0.1066	0.93	1000	0.1097	11.4191
0.0707	1.4	1500	0.0948	10.0972
0.0649	1.87	2000	0.0824	8.7874
0.0249	2.34	2500	0.0828	8.6930
0.0275	2.8	3000	0.0792	7.8402
0.0139	3.27	3500	0.0748	6.7619
0.0121	3.74	4000	0.0766	7.2492
0.0071	4.21	4500	0.0759	6.5335
0.005	4.67	5000	0.0764	6.3903
0.0036	5.14	5500	0.0768	6.0217
0.0037	5.61	6000	0.0770	6.1009
0.0013	6.07	6500	0.0768	5.9182
0.0012	6.54	7000	0.0765	5.7933
0.0014	7.01	7500	0.0770	5.8299
0.0008	7.48	8000	0.0777	5.7842

Framework versions

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