rlimonta/wav2vec2-large-xlsr-53-portuguese

Wav2Vec2 Large XLSR-53 Portuguese - Ricardo Limonta

This model is a fine-tuned version of lgris/wav2vec2-large-xlsr-open-brazilian-portuguese-v2 on the Common Voice 13 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1656
• Wer: 0.1209

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: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 8
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
8.0499	0.6361	1000	0.2436	0.1808
0.4775	1.2723	2000	0.2135	0.1544
0.4073	1.9084	3000	0.1852	0.1528
0.3635	2.5445	4000	0.1902	0.1451
0.3448	3.1807	5000	0.1859	0.1403
0.3195	3.8168	6000	0.1932	0.1379
0.2912	4.4529	7000	0.1799	0.1327
0.2714	5.0891	8000	0.1746	0.1277
0.2598	5.7252	9000	0.1794	0.1258
0.2477	6.3613	10000	0.1706	0.1233
0.2347	6.9975	11000	0.1682	0.1213
0.2235	7.6336	12000	0.1656	0.1209

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1