xtreme_s_xlsr_300m_voxpopuli_en

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the GOOGLE/XTREME_S - VOXPOPULI.EN dataset. It achieves the following results on the evaluation set:

  • Cer: 0.0966
  • Loss: 0.3127
  • Wer: 0.1549
  • Predict Samples: 1842

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: 1
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 8
  • total_train_batch_size: 64
  • total_eval_batch_size: 8
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 2000
  • num_epochs: 10.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
1.4221 0.19 500 1.3325 0.8224 0.3432
0.8429 0.38 1000 0.7087 0.5028 0.2023
0.7377 0.57 1500 0.4900 0.2778 0.1339
0.5641 0.77 2000 0.4460 0.2540 0.1284
0.5787 0.96 2500 0.4242 0.2148 0.1167
0.3465 1.15 3000 0.4210 0.2087 0.1154
0.2787 1.34 3500 0.3954 0.2090 0.1155
0.2775 1.53 4000 0.3938 0.1992 0.1133
0.262 1.72 4500 0.3748 0.2104 0.1151
0.3138 1.92 5000 0.3825 0.1993 0.1134
0.4331 2.11 5500 0.3648 0.1935 0.1104
0.3802 2.3 6000 0.3966 0.1910 0.1109
0.3928 2.49 6500 0.3995 0.1898 0.1100
0.3441 2.68 7000 0.3764 0.1887 0.1103
0.3673 2.87 7500 0.3800 0.1843 0.1086
0.3422 3.07 8000 0.3932 0.1830 0.1092
0.2933 3.26 8500 0.3672 0.1915 0.1104
0.1785 3.45 9000 0.3820 0.1796 0.1072
0.321 3.64 9500 0.3533 0.1994 0.1126
0.1673 3.83 10000 0.3683 0.1856 0.1084
0.1757 4.02 10500 0.3365 0.1925 0.1102
0.1881 4.22 11000 0.3528 0.1775 0.1066
0.3106 4.41 11500 0.3909 0.1754 0.1063
0.25 4.6 12000 0.3734 0.1723 0.1052
0.2005 4.79 12500 0.3358 0.1900 0.1092
0.2982 4.98 13000 0.3513 0.1766 0.1060
0.1552 5.17 13500 0.3720 0.1729 0.1059
0.1645 5.37 14000 0.3569 0.1713 0.1044
0.2065 5.56 14500 0.3639 0.1720 0.1048
0.1898 5.75 15000 0.3660 0.1726 0.1050
0.1397 5.94 15500 0.3731 0.1670 0.1033
0.2056 6.13 16000 0.3782 0.1650 0.1030
0.1859 6.32 16500 0.3903 0.1667 0.1033
0.1374 6.52 17000 0.3721 0.1736 0.1048
0.2482 6.71 17500 0.3899 0.1643 0.1023
0.159 6.9 18000 0.3847 0.1687 0.1032
0.1487 7.09 18500 0.3817 0.1671 0.1030
0.1942 7.28 19000 0.4120 0.1616 0.1018
0.1517 7.47 19500 0.3856 0.1635 0.1020
0.0946 7.67 20000 0.3838 0.1621 0.1016
0.1455 7.86 20500 0.3749 0.1652 0.1020
0.1303 8.05 21000 0.4074 0.1615 0.1011
0.1207 8.24 21500 0.4121 0.1606 0.1008
0.0727 8.43 22000 0.3948 0.1607 0.1009
0.1123 8.62 22500 0.4025 0.1603 0.1009
0.1606 8.82 23000 0.3963 0.1580 0.1004
0.1458 9.01 23500 0.3991 0.1574 0.1002
0.2286 9.2 24000 0.4149 0.1596 0.1009
0.1284 9.39 24500 0.4251 0.1572 0.1002
0.1141 9.58 25000 0.4264 0.1579 0.1002
0.1823 9.77 25500 0.4230 0.1562 0.0999
0.2514 9.97 26000 0.4242 0.1564 0.0999

Framework versions

  • Transformers 4.18.0.dev0
  • Pytorch 1.10.1+cu111
  • Datasets 1.18.4.dev0
  • Tokenizers 0.11.6
Downloads last month
16
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 anton-l/xtreme_s_xlsr_300m_voxpopuli_en