wav2vec2_l2arctic

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.5487
• Wer: 0.1460
• Cer: 0.0904

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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
6.3356	0.9941	84	4.6295	1.0	1.0
3.5484	2.0	169	3.5272	1.0	1.0
3.5314	2.9941	253	3.5110	1.0	1.0
3.5084	4.0	338	3.5019	1.0	1.0
3.3271	4.9941	422	3.2417	1.0	1.0
1.6302	6.0	507	1.0777	0.3444	0.3220
0.7834	6.9941	591	0.6123	0.1780	0.1189
0.6067	8.0	676	0.5169	0.1550	0.0983
0.534	8.9941	760	0.5095	0.1549	0.0993
0.4711	10.0	845	0.4976	0.1524	0.0962
0.3979	10.9941	929	0.4951	0.1497	0.0937
0.354	12.0	1014	0.5012	0.1505	0.0943
0.3415	12.9941	1098	0.5090	0.1489	0.0937
0.295	14.0	1183	0.5098	0.1488	0.0944
0.2917	14.9941	1267	0.5296	0.1507	0.0946
0.2397	16.0	1352	0.5315	0.1507	0.0944
0.2713	16.9941	1436	0.5367	0.1467	0.0913
0.2153	18.0	1521	0.5456	0.1483	0.0924
0.206	18.9941	1605	0.5464	0.1471	0.0914
0.2488	19.8817	1680	0.5487	0.1460	0.0904

Framework versions

• Transformers 4.40.0
• Pytorch 2.2.1+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1