khalidey's picture
Update README.md
e5fb328
metadata
license: apache-2.0

Whisper Small - Swedish

This is a fine-tuned version of the openai/whisper-small model on the Common Voice 11.0 dataset.

The following results were achieved after training for 4000 optimization steps:

  • Training Loss: 0.003900
  • Validation Loss: 0.326255
  • WER: 19.894598

Training hyperparameters

The following hyperparameters were used during training:

  • train_batch_size: 16
  • gradient_accumulation_steps: 1
  • learning_rate: 1e-5
  • eval_batch_size: 8
  • max_steps: 4000
  • eval_steps: 1000

Framework Versions

  • Transformers 4.25.0
  • Pytorch 1.12.1
  • Datasets 2.7.1