VladS159/Whisper_medium_ro_VladS_2500_steps_multi_gpu_26_02_2024

Whisper Medium Ro - Sarbu Vlad - multi gpu

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: 0.1521
• Wer: 11.9574

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: 250
• training_steps: 2500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.1596	0.61	250	0.1685	15.0114
0.0731	1.23	500	0.1366	12.9643
0.0737	1.84	750	0.1317	12.7574
0.032	2.46	1000	0.1292	11.8935
0.0219	3.07	1250	0.1296	11.9635
0.0169	3.69	1500	0.1359	11.9179
0.0106	4.3	1750	0.1422	12.0608
0.0103	4.91	2000	0.1434	11.8357
0.0058	5.53	2250	0.1499	11.8144
0.0047	6.14	2500	0.1521	11.9574

Framework versions

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