VladS159/Whisper_medium_ro_VladS_10000_steps_multi-gpu_smaller_lr

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

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

• Loss: 0.1249
• Wer: 10.5571

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: 16
• eval_batch_size: 10
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• total_train_batch_size: 48
• 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: 600
• training_steps: 6000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2432	0.68	500	0.2134	19.7435
0.137	1.36	1000	0.1532	15.5189
0.0672	2.04	1500	0.1287	13.0426
0.0579	2.72	2000	0.1218	12.8659
0.0307	3.4	2500	0.1183	11.9887
0.0167	4.08	3000	0.1177	11.5866
0.016	4.76	3500	0.1149	10.9531
0.0099	5.43	4000	0.1212	10.9713
0.0058	6.11	4500	0.1216	10.8251
0.0056	6.79	5000	0.1224	10.6515
0.0036	7.47	5500	0.1238	10.6211
0.0035	8.15	6000	0.1249	10.5571

Framework versions

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