wav2vec2-lar-xlsr-es-col

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

• Loss: 0.0947
• Wer: 0.1884

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
4.8446	8.51	400	2.8174	0.9854
0.5146	17.02	800	0.1022	0.2020
0.0706	25.53	1200	0.0947	0.1884

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.1+cu102
• Datasets 1.13.3
• Tokenizers 0.10.3