Edit model card

wav2vec2-xls-r-300m-uk

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset.
Notebook for training is located in this repository: https://github.com/robinhad/wav2vec2-xls-r-ukrainian.
It achieves the following results on the evaluation set:

  • Loss: 0.4165
  • Wer: 0.2799
  • Cer: 0.0601

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: 0.0003
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 20
  • total_train_batch_size: 160
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 500
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Cer Validation Loss Wer
4.3982 9.3 400 0.1437 0.5218 0.6507
0.229 18.6 800 0.0848 0.3679 0.4048
0.1054 27.9 1200 0.0778 0.3813 0.3670
0.0784 37.21 1600 0.0747 0.3839 0.3550
0.066 46.51 2000 0.0736 0.3970 0.3443
0.0603 55.8 2400 0.0722 0.3702 0.3393
0.0539 65.11 2800 0.0724 0.3762 0.3388
0.0497 74.41 3200 0.0713 0.3623 0.3414
0.0432 83.71 3600 0.0725 0.3847 0.3346
0.0438 93.02 4000 0.0750 0.4058 0.3393
0.0413 102.32 4400 0.0727 0.3957 0.3363
0.039 111.62 4800 0.0718 0.3865 0.3330
0.0356 120.92 5200 0.0711 0.3860 0.3319
0.0336 130.23 5600 0.0700 0.3902 0.3242
0.034 139.53 6000 0.0732 0.3930 0.3337
0.0273 148.83 6400 0.0748 0.3912 0.3375
0.027 158.14 6800 0.0752 0.4266 0.3434
0.028 167.44 7200 0.0708 0.3895 0.3227
0.0241 176.73 7600 0.0727 0.3967 0.3294
0.0241 186.05 8000 0.0712 0.4058 0.3255
0.0209 195.34 8400 0.0702 0.4102 0.3233
0.0206 204.64 8800 0.0699 0.4075 0.3194
0.0172 213.94 9200 0.0695 0.4222 0.3191
0.0166 223.25 9600 0.0678 0.3860 0.3135
0.0156 232.55 10000 0.0677 0.4035 0.3117
0.0149 241.85 10400 0.0677 0.3951 0.3087
0.0142 251.16 10800 0.0674 0.3972 0.3097
0.0134 260.46 11200 0.0675 0.4069 0.3111
0.0116 269.76 11600 0.0697 0.4189 0.3161
0.0119 279.07 12000 0.0648 0.3902 0.3008
0.0098 288.37 12400 0.0652 0.4095 0.3002
0.0091 297.67 12800 0.0644 0.3892 0.2990
0.0094 306.96 13200 0.0647 0.4026 0.2983
0.0081 316.28 13600 0.0646 0.4303 0.2978
0.0079 325.57 14000 0.0643 0.4044 0.2980
0.0072 334.87 14400 0.0655 0.3828 0.2999
0.0081 344.18 14800 0.0668 0.4108 0.3046
0.0088 353.48 15200 0.0654 0.4019 0.2993
0.0088 362.78 15600 0.0681 0.4073 0.3091
0.0079 372.09 16000 0.0667 0.4204 0.3055
0.0072 381.39 16400 0.0656 0.4030 0.3028
0.0073 390.69 16800 0.0677 0.4032 0.3081
0.0069 399.99 17200 0.0669 0.4130 0.3021
0.0063 409.3 17600 0.0651 0.4072 0.2979
0.0059 418.6 18000 0.0640 0.4110 0.2969
0.0056 427.9 18400 0.0647 0.4229 0.2995
0.005 437.21 18800 0.0624 0.4118 0.2885
0.0046 446.51 19200 0.0615 0.4111 0.2841
0.0043 455.8 19600 0.0616 0.4071 0.2850
0.0038 465.11 20000 0.0624 0.4268 0.2867
0.0035 474.41 20400 0.0605 0.4117 0.2820
0.0035 483.71 20800 0.0602 0.4155 0.2819
0.0034 493.02 21200 0.0601 0.4165 0.2799

Framework versions

  • Transformers 4.14.1
  • Pytorch 1.10.0
  • Datasets 1.16.1
  • Tokenizers 0.10.3
Downloads last month
32
Hosted inference API
or or
This model can be loaded on the Inference API on-demand.

Dataset used to train robinhad/wav2vec2-xls-r-300m-uk

Space using robinhad/wav2vec2-xls-r-300m-uk 1

Evaluation results