wav2vec2-1b-E30_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: 0.4977
• Cer: 13.7277

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
13.4278	0.2580	200	4.0083	87.1123
2.1559	0.5160	400	1.8970	40.9833
1.3277	0.7741	600	1.2101	31.0620
1.162	1.0321	800	1.0824	26.5096
0.9949	1.2901	1000	0.9657	24.2246
0.9109	1.5481	1200	1.0152	24.8414
0.8943	1.8062	1400	0.8544	21.7869
0.7895	2.0642	1600	0.9202	22.9617
0.6679	2.3222	1800	0.9574	24.1835
0.6296	2.5802	2000	0.7541	19.2199
0.6245	2.8383	2200	0.7259	19.2728
0.5656	3.0963	2400	0.6447	17.3344
0.4821	3.3543	2600	0.6489	16.9878
0.4513	3.6123	2800	0.6556	17.5282
0.4285	3.8703	3000	0.6180	16.7234
0.374	4.1284	3200	0.5651	15.2314
0.3375	4.3864	3400	0.5135	13.8275
0.3158	4.6444	3600	0.4945	13.7688
0.2897	4.9024	3800	0.4977	13.7277

Framework versions

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