wav2vec2-large-xls-r-300m-nl

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the test set:

  • Loss: 0.3923
  • Wer: 0.1748

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: 7.5e-05
  • 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: 500
  • num_epochs: 50
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
1.5787 0.89 400 0.6354 0.5643
0.3036 1.78 800 0.3690 0.3552
0.188 2.67 1200 0.3239 0.2958
0.1434 3.56 1600 0.3093 0.2515
0.1245 4.44 2000 0.3024 0.2433
0.1095 5.33 2400 0.3249 0.2643
0.0979 6.22 2800 0.3191 0.2281
0.0915 7.11 3200 0.3152 0.2216
0.0829 8.0 3600 0.3419 0.2218
0.0777 8.89 4000 0.3432 0.2132
0.073 9.78 4400 0.3223 0.2131
0.0688 10.67 4800 0.3094 0.2152
0.0647 11.56 5200 0.3411 0.2152
0.0639 12.44 5600 0.3762 0.2135
0.0599 13.33 6000 0.3790 0.2137
0.0572 14.22 6400 0.3693 0.2118
0.0563 15.11 6800 0.3495 0.2139
0.0521 16.0 7200 0.3800 0.2023
0.0508 16.89 7600 0.3678 0.2033
0.0513 17.78 8000 0.3845 0.1987
0.0476 18.67 8400 0.3511 0.2037
0.045 19.56 8800 0.3794 0.1994
0.044 20.44 9200 0.3525 0.2050
0.043 21.33 9600 0.4082 0.2007
0.0409 22.22 10000 0.3866 0.2004
0.0393 23.11 10400 0.3899 0.2008
0.0382 24.0 10800 0.3626 0.1951
0.039 24.89 11200 0.3936 0.1953
0.0361 25.78 11600 0.4262 0.1928
0.0362 26.67 12000 0.3796 0.1934
0.033 27.56 12400 0.3616 0.1934
0.0321 28.44 12800 0.3742 0.1933
0.0325 29.33 13200 0.3582 0.1869
0.0309 30.22 13600 0.3717 0.1874
0.029 31.11 14000 0.3814 0.1894
0.0296 32.0 14400 0.3698 0.1877
0.0281 32.89 14800 0.3976 0.1899
0.0275 33.78 15200 0.3854 0.1858
0.0264 34.67 15600 0.4021 0.1889
0.0261 35.56 16000 0.3850 0.1830
0.0242 36.44 16400 0.4091 0.1878
0.0245 37.33 16800 0.4012 0.1846
0.0243 38.22 17200 0.3996 0.1833
0.0223 39.11 17600 0.3962 0.1815
0.0223 40.0 18000 0.3898 0.1832
0.0219 40.89 18400 0.4019 0.1822
0.0211 41.78 18800 0.4035 0.1809
0.021 42.67 19200 0.3915 0.1826
0.0208 43.56 19600 0.3934 0.1784
0.0188 44.44 20000 0.3912 0.1787
0.0195 45.33 20400 0.3989 0.1766
0.0186 46.22 20800 0.3887 0.1773
0.0188 47.11 21200 0.3982 0.1758
0.0175 48.0 21600 0.3933 0.1755
0.0172 48.89 22000 0.3921 0.1749
0.0187 49.78 22400 0.3923 0.1748

Framework versions

  • Transformers 4.16.0.dev0
  • Pytorch 1.10.1+cu102
  • Datasets 1.17.1.dev0
  • Tokenizers 0.11.0
Downloads last month
14
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train RuudVelo/wav2vec2-large-xls-r-300m-nl

Evaluation results