Edit model card

wav2vec2-xls-r-1b-dv

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the common_voice dataset. It achieves the following results on the evaluation set:

  • Loss: 0.1702
  • Wer: 0.2123

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: 4.5e-05
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 4
  • 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: 500
  • num_epochs: 30
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
3.8412 0.66 400 0.7160 0.7913
0.6832 1.33 800 0.3401 0.5268
0.4624 1.99 1200 0.2671 0.4683
0.3832 2.65 1600 0.2395 0.4410
0.3443 3.32 2000 0.2410 0.4296
0.324 3.98 2400 0.2302 0.4143
0.2934 4.64 2800 0.2402 0.4136
0.2773 5.31 3200 0.2134 0.4088
0.2638 5.97 3600 0.2072 0.4037
0.2479 6.63 4000 0.2036 0.3876
0.2424 7.3 4400 0.2037 0.3767
0.2249 7.96 4800 0.1959 0.3802
0.2169 8.62 5200 0.1943 0.3813
0.2109 9.29 5600 0.1944 0.3691
0.1991 9.95 6000 0.1870 0.3589
0.1917 10.61 6400 0.1834 0.3485
0.1862 11.28 6800 0.1857 0.3486
0.1744 11.94 7200 0.1812 0.3330
0.171 12.6 7600 0.1797 0.3436
0.1599 13.27 8000 0.1839 0.3319
0.1597 13.93 8400 0.1737 0.3385
0.1494 14.59 8800 0.1807 0.3239
0.1444 15.26 9200 0.1750 0.3155
0.1382 15.92 9600 0.1705 0.3084
0.1299 16.58 10000 0.1777 0.2999
0.1306 17.25 10400 0.1765 0.3056
0.1239 17.91 10800 0.1676 0.2864
0.1149 18.57 11200 0.1774 0.2861
0.1134 19.24 11600 0.1654 0.2699
0.1101 19.9 12000 0.1621 0.2651
0.1038 20.56 12400 0.1686 0.2610
0.1038 21.23 12800 0.1722 0.2559
0.0988 21.89 13200 0.1708 0.2486
0.0949 22.55 13600 0.1696 0.2453
0.0913 23.22 14000 0.1677 0.2424
0.0879 23.88 14400 0.1640 0.2359
0.0888 24.54 14800 0.1697 0.2347
0.0826 25.21 15200 0.1709 0.2314
0.0819 25.87 15600 0.1679 0.2256
0.0793 26.53 16000 0.1701 0.2214
0.0773 27.2 16400 0.1682 0.2176
0.0783 27.86 16800 0.1685 0.2165
0.074 28.52 17200 0.1688 0.2155
0.0753 29.19 17600 0.1695 0.2110
0.0699 29.85 18000 0.1702 0.2123

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.3
  • Tokenizers 0.11.0
Downloads last month
23

Dataset used to train shahukareem/wav2vec2-xls-r-1b-dv

Evaluation results