wav2vec2-1b-E10_freq

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: 0.5460
• Cer: 14.6793

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
9.6681	0.2580	200	3.0566	61.6424
1.8046	0.5160	400	2.0731	44.9601
1.3046	0.7741	600	1.4756	36.8832
1.1537	1.0321	800	1.3063	32.3484
0.9218	1.2901	1000	1.0777	26.4215
0.8464	1.5481	1200	1.0670	27.6081
0.7727	1.8062	1400	0.9148	23.2906
0.6984	2.0642	1600	0.9548	24.3832
0.5856	2.3222	1800	0.9348	23.2965
0.5365	2.5802	2000	0.8671	22.4213
0.5223	2.8383	2200	0.8267	20.7766
0.4623	3.0963	2400	0.7185	18.1450
0.3774	3.3543	2600	0.7324	19.4784
0.3467	3.6123	2800	0.6037	15.8012
0.3195	3.8703	3000	0.6357	16.4767
0.2759	4.1284	3200	0.6200	16.5237
0.2241	4.3864	3400	0.5901	15.4781
0.2211	4.6444	3600	0.5591	14.9436
0.2087	4.9024	3800	0.5460	14.6793

Framework versions

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