Edit model card

dat259-wav2vec2-en2-copy

This model is a fine-tuned version of facebook/wav2vec2-base on the common_voice_1_0 dataset. It achieves the following results on the evaluation set:

  • Loss: 1.5429
  • Wer: 0.5569

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: 0.0001
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 300
  • num_epochs: 20
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
4.5149 1.82 200 3.0474 1.0
2.0483 3.64 400 1.8004 0.7808
0.5464 5.45 600 1.5242 0.6529
0.3269 7.27 800 1.3820 0.6046
0.2337 9.09 1000 1.4605 0.5866
0.1863 10.91 1200 1.5567 0.5864
0.1531 12.73 1400 1.5787 0.5871
0.1326 14.55 1600 1.5766 0.5679
0.1118 16.36 1800 1.4709 0.5566
0.1017 18.18 2000 1.5455 0.5598
0.0955 20.0 2200 1.5429 0.5569

Framework versions

  • Transformers 4.21.1
  • Pytorch 1.12.1+cu102
  • Datasets 2.4.0
  • Tokenizers 0.12.1
Downloads last month
1