wav2vec2-xls-r-300m-Turkish-Tr-med

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 evaluation set:

  • Loss: 0.4727
  • Wer: 0.4677

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: 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: 60
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
4.8093 4.21 400 2.7831 1.0
0.9881 8.42 800 0.5088 0.6681
0.3519 12.63 1200 0.4496 0.6007
0.2436 16.84 1600 0.4993 0.5654
0.1874 21.05 2000 0.4793 0.5530
0.1561 25.26 2400 0.5187 0.5589
0.1336 29.47 2800 0.5135 0.5311
0.1163 33.68 3200 0.4960 0.5143
0.1056 37.89 3600 0.4795 0.5045
0.0959 42.11 4000 0.4883 0.4987
0.0819 46.32 4400 0.4799 0.4903
0.0756 50.53 4800 0.4822 0.4831
0.0692 54.74 5200 0.4621 0.4762
0.062 58.95 5600 0.4727 0.4677

Framework versions

  • Transformers 4.11.3
  • Pytorch 1.10.0+cu111
  • Datasets 1.14.0
  • Tokenizers 0.10.3
Downloads last month
10
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 emre/wav2vec2-xls-r-300m-Turkish-Tr-med