wav2vec2-large-xls-r-1b-irish-colab

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

• Loss: 1.0795
• Wer: 46.91

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: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.6902	12.12	400	1.1158	0.5959
0.2988	24.24	800	1.1375	0.5094

Framework versions

• Transformers 4.25.1
• Pytorch 1.10.0+cu113
• Datasets 2.0.0
• Tokenizers 0.13.2