Edit model card

wav2vec2-large-xlsr-korean-demo-with-LM

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:

  • Loss: 0.3015
  • Wer: 0.2113

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: 4
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 8
  • 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
4.7496 1.08 400 3.1801 1.0
1.4505 2.16 800 0.5090 0.5659
0.566 3.23 1200 0.3600 0.4039
0.4265 4.31 1600 0.3224 0.3639
0.3611 5.39 2000 0.3152 0.3575
0.3035 6.47 2400 0.2814 0.3054
0.2863 7.55 2800 0.2749 0.2923
0.247 8.63 3200 0.2787 0.2884
0.232 9.7 3600 0.2924 0.2788
0.2069 10.78 4000 0.2668 0.2694
0.1922 11.86 4400 0.2873 0.2667
0.1747 12.94 4800 0.2870 0.2589
0.1755 14.02 5200 0.2778 0.2543
0.1546 15.09 5600 0.3062 0.2621
0.1456 16.17 6000 0.3043 0.2479
0.1404 17.25 6400 0.2885 0.2443
0.1308 18.33 6800 0.3274 0.2417
0.125 19.41 7200 0.2922 0.2401
0.1148 20.49 7600 0.2899 0.2300
0.1129 21.56 8000 0.2963 0.2276
0.1086 22.64 8400 0.2903 0.2209
0.097 23.72 8800 0.3041 0.2220
0.099 24.8 9200 0.2870 0.2168
0.0905 25.88 9600 0.2992 0.2176
0.0929 26.95 10000 0.2934 0.2115
0.0827 28.03 10400 0.2945 0.2141
0.0818 29.11 10800 0.3015 0.2113

Usage

Framework versions

  • Transformers 4.21.0
  • Pytorch 1.12.0+cu113
  • Datasets 2.4.0
  • Tokenizers 0.12.1
Downloads last month
11
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.