whisper-dpv-finetuned-WITH-AUGMENTATION-LOWER-LR

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

  • Loss: 0.5717
  • Wer: 34.5241

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-06
  • train_batch_size: 1
  • eval_batch_size: 1
  • seed: 42
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 4
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 100
  • num_epochs: 4
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
0.6221 0.62 1000 0.5345 35.9711
0.4318 1.25 2000 0.5271 34.9537
0.3859 1.87 3000 0.5338 34.3658
0.3005 2.49 4000 0.5532 34.8858
0.2444 3.12 5000 0.5628 33.7102
0.315 3.74 6000 0.5717 34.5241

Framework versions

  • Transformers 4.26.0.dev0
  • Pytorch 1.12.1
  • Datasets 2.7.1
  • Tokenizers 0.13.2
Downloads last month
2
Inference Providers NEW
This model is not currently available via any of the supported third-party Inference Providers, and the model is not deployed on the HF Inference API.