whisper-large-final

This model is a fine-tuned version of openai/whisper-large-v2 on an unknown dataset. It achieves the following results on the evaluation set:

• eval_loss: 0.0112
• eval_wer: 1.1712
• eval_runtime: 982.7637
• eval_samples_per_second: 1.892
• eval_steps_per_second: 0.237
• epoch: 6.4205
• step: 4000

Model description

Step Training Loss Validation Loss Wer 500 0.431500 0.412413 48.265244 1000 0.244500 0.230148 29.284654 1500 0.134300 0.122366 16.588772 2000 0.055800 0.069241 10.551493 2500 0.045700 0.035967 4.860615 3000 0.027900 0.024117 3.425524 3500 0.011000 0.016053 1.770495 4000 0.004800 0.011227 1.171166

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: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• training_steps: 5000
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.40.2
• Pytorch 2.2.0
• Datasets 2.19.1
• Tokenizers 0.19.1