wav2vec2-1b-E30_freq_pause_speed

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: 0.5381
• Cer: 14.3033

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
9.3347	0.2580	200	3.0510	67.1170
1.6722	0.5160	400	1.7179	39.9495
1.1366	0.7741	600	1.4161	34.0343
0.9454	1.0321	800	1.0728	26.3217
0.8179	1.2901	1000	1.0363	26.1219
0.7504	1.5481	1200	0.9902	26.1748
0.676	1.8062	1400	0.9778	24.4008
0.6309	2.0642	1600	0.9562	24.5359
0.523	2.3222	1800	0.9541	24.5888
0.469	2.5802	2000	0.8640	22.1041
0.4512	2.8383	2200	0.7965	20.3595
0.3876	3.0963	2400	0.6971	18.5914
0.3249	3.3543	2600	0.6755	18.1450
0.2832	3.6123	2800	0.6407	16.8174
0.2836	3.8703	3000	0.6073	16.1125
0.2379	4.1284	3200	0.6392	16.3886
0.1937	4.3864	3400	0.5707	14.9965
0.1871	4.6444	3600	0.5454	14.7263
0.1713	4.9024	3800	0.5381	14.3033

Framework versions

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