Edit model card

wav2vec2-large-xls-r-300m-hungarian

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HU dataset. It achieves the following results on the evaluation set:

  • Loss: 0.2562
  • Wer: 0.3112

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: 7e-05
  • train_batch_size: 32
  • eval_batch_size: 32
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 1000
  • num_epochs: 50.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.3964 3.52 1000 1.2251 0.8781
1.3176 7.04 2000 0.3872 0.4462
1.1999 10.56 3000 0.3244 0.3922
1.1633 14.08 4000 0.3014 0.3704
1.1132 17.61 5000 0.2913 0.3623
1.0888 21.13 6000 0.2864 0.3498
1.0487 24.65 7000 0.2821 0.3435
1.0431 28.17 8000 0.2739 0.3308
0.9896 31.69 9000 0.2629 0.3243
0.9839 35.21 10000 0.2806 0.3308
0.9586 38.73 11000 0.2650 0.3235
0.9501 42.25 12000 0.2585 0.3173
0.938 45.77 13000 0.2561 0.3117
0.921 49.3 14000 0.2559 0.3115

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0
Downloads last month
10

Dataset used to train infinitejoy/wav2vec2-large-xls-r-300m-hungarian

Evaluation results