wav2vec2-1b-Y2

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: 0.9063
• Cer: 22.3214

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
14.1704	0.2580	200	4.3519	97.8090
2.4465	0.5160	400	2.2289	45.2126
1.3213	0.7741	600	1.7263	39.3386
1.0017	1.0321	800	1.2111	29.5348
0.8038	1.2901	1000	1.4298	34.3339
0.7428	1.5481	1200	1.3785	32.6656
0.705	1.8062	1400	1.2020	30.1104
0.5742	2.0642	1600	1.1142	27.8959
0.4805	2.3222	1800	1.3923	33.2237
0.471	2.5802	2000	1.1341	27.9840
0.4327	2.8383	2200	0.9932	25.4582
0.3677	3.0963	2400	0.8940	22.6739
0.2976	3.3543	2600	0.9628	24.6182
0.2802	3.6123	2800	0.8832	22.1570
0.2637	3.8703	3000	0.9087	22.8207
0.2175	4.1284	3200	0.9277	22.7326
0.1932	4.3864	3400	0.9152	22.6328
0.1724	4.6444	3600	0.9290	22.6621
0.1693	4.9024	3800	0.9063	22.3214

Framework versions

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