cminja/whisper-tiny-sr-commonvoice-2000

finetuned whisper-tiny model on custom dataset

This model is a fine-tuned version of openai/whisper-tiny on Serbian Mozilla/Common Voice 13. It achieves the following results on the evaluation set:

• Loss: 0.1628
• Wer Ortho: 0.1635
• Wer: 0.0556

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-5
• train_batch_size: 32
• eval_batch_size: 32
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• training_steps: 2000
• mixed_precision_training: Native AMP

Training Results

Training Loss	Epoch	Step	Validation Loss	Wer Ortho	Wer
0.0600	1.34	500	0.1852	0.1800	0.0745
0.0285	2.67	1000	0.1715	0.1710	0.0640
0.0140	4.01	1500	0.1658	0.1685	0.0582

Framework Versions

• Transformers: 4.41.2
• Pytorch: 2.3.0+cu121
• Datasets: 2.18.0
• Tokenizers: 0.19.1