dat259-wav2vec2-en

This model is a fine-tuned version of facebook/wav2vec2-base on the common_voice_1_0 dataset. It achieves the following results on the evaluation set:

• Loss: 1.5042
• Wer: 0.5793

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: 16
• 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: 200
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.2531	1.82	200	3.0566	1.0
2.1194	3.64	400	1.6370	0.7706
0.6464	5.45	600	1.3950	0.6694
0.3891	7.27	800	1.4443	0.6525
0.2783	9.09	1000	1.4309	0.6152
0.2088	10.91	1200	1.3592	0.5960
0.1685	12.73	1400	1.4690	0.6031
0.1397	14.55	1600	1.4691	0.5819
0.1209	16.36	1800	1.5004	0.5840
0.1122	18.18	2000	1.5069	0.5806
0.1025	20.0	2200	1.5042	0.5793

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.1+cu102
• Datasets 2.4.0
• Tokenizers 0.12.1