wav2vec2-1b-E50_speed2

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.7405
• Cer: 24.6417

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: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
26.7947	0.2580	200	30.8746	111.8127
7.993	0.5160	400	5.7855	94.2317
4.7455	0.7741	600	4.5989	93.9262
4.4924	1.0321	800	4.7152	93.2331
4.3711	1.2901	1000	4.7883	93.3564
4.3243	1.5481	1200	4.7205	92.4636
4.2683	1.8062	1400	4.5020	90.9305
4.1967	2.0642	1600	4.8431	92.8102
4.0612	2.3222	1800	4.8480	92.2697
3.6744	2.5802	2000	3.8378	77.9664
2.8934	2.8383	2200	2.9635	61.9713
2.1459	3.0963	2400	2.1180	51.9267
1.5896	3.3543	2600	1.5030	38.3870
1.0885	3.6123	2800	1.1457	30.6978
0.8797	3.8703	3000	0.9902	29.2234
0.7317	4.1284	3200	0.8634	25.7636
0.6046	4.3864	3400	0.7911	24.6476
0.5618	4.6444	3600	0.7773	25.6285
0.5027	4.9024	3800	0.7405	24.6417

Framework versions

• Transformers 4.45.2
• Pytorch 2.3.1.post100
• Datasets 2.19.1
• Tokenizers 0.20.1