wav2vec2-large-xls-r-53m-gl-jupyter2

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

• Loss: 0.0941
• Wer: 0.0615

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: 45
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.7298	3.36	400	0.2477	0.2493
0.1507	6.72	800	0.1294	0.1264
0.066	10.08	1200	0.1235	0.1161
0.0456	13.44	1600	0.1011	0.1001
0.0347	16.8	2000	0.1033	0.0909
0.0284	20.17	2400	0.1083	0.0861
0.0221	23.53	2800	0.1010	0.0761
0.0199	26.89	3200	0.0911	0.0754
0.0155	30.25	3600	0.1026	0.0743
0.0142	33.61	4000	0.1024	0.0719
0.0125	36.97	4400	0.0977	0.0676
0.0104	40.33	4800	0.0945	0.0664
0.0089	43.69	5200	0.0941	0.0615

Framework versions

• Transformers 4.11.3
• Pytorch 1.9.1+cu111
• Datasets 1.18.3
• Tokenizers 0.10.3