ls-timit-wsj0-100percent-supervised-aug

This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0489
• Wer: 0.0275

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: 0.0001
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3491	4.57	1000	0.0470	0.0416
0.1088	9.13	2000	0.0582	0.0343
0.0702	13.7	3000	0.0471	0.0271
0.0532	18.26	4000	0.0489	0.0275

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.2
• Datasets 1.18.2
• Tokenizers 0.10.3