Edit model card

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - HY-AM dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4521
  • Wer: 0.5141
  • Cer: 0.1100
  • Wer+LM: 0.2756
  • Cer+LM: 0.0866

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: 8e-05
  • train_batch_size: 16
  • eval_batch_size: 64
  • seed: 42
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 128
  • optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
  • lr_scheduler_type: tristage
  • lr_scheduler_ratios: [0.1, 0.4, 0.5]
  • training_steps: 1400
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer Cer
6.1298 19.87 100 3.1204 1.0 1.0
2.7269 39.87 200 0.6200 0.7592 0.1755
1.4643 59.87 300 0.4796 0.5921 0.1277
1.1242 79.87 400 0.4637 0.5359 0.1145
0.9592 99.87 500 0.4521 0.5141 0.1100
0.8704 119.87 600 0.4736 0.4914 0.1045
0.7908 139.87 700 0.5394 0.5250 0.1124
0.7049 159.87 800 0.4822 0.4754 0.0985
0.6299 179.87 900 0.4890 0.4809 0.1028
0.5832 199.87 1000 0.5233 0.4813 0.1028
0.5145 219.87 1100 0.5350 0.4781 0.0994
0.4604 239.87 1200 0.5223 0.4715 0.0984
0.4226 259.87 1300 0.5167 0.4625 0.0953
0.3946 279.87 1400 0.5248 0.4614 0.0950

Framework versions

  • Transformers 4.17.0.dev0
  • Pytorch 1.10.2+cu102
  • Datasets 1.18.2.dev0
  • Tokenizers 0.11.0
Downloads last month
24
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 arampacha/wav2vec2-xls-r-1b-hy-cv

Evaluation results