wav2vec2-l-xlsr-es-col-pro-noise

This model is a fine-tuned version of jonatasgrosman/wav2vec2-large-xlsr-53-spanish on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0677
• Wer: 0.0380

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.94	1.21	400	0.0800	0.0814
0.4711	2.42	800	0.0730	0.0692
0.3451	3.62	1200	0.0729	0.0669
0.2958	4.83	1600	0.0796	0.0667
0.2544	6.04	2000	0.0808	0.0584
0.227	7.25	2400	0.0791	0.0643
0.2061	8.46	2800	0.0718	0.0582
0.1901	9.67	3200	0.0709	0.0587
0.179	10.87	3600	0.0698	0.0558
0.1693	12.08	4000	0.0709	0.0530
0.1621	13.29	4400	0.0640	0.0487
0.1443	14.5	4800	0.0793	0.0587
0.1408	15.71	5200	0.0741	0.0528
0.1377	16.92	5600	0.0702	0.0462
0.1292	18.13	6000	0.0822	0.0539
0.1197	19.33	6400	0.0625	0.0436
0.1137	20.54	6800	0.0650	0.0419
0.1017	21.75	7200	0.0630	0.0392
0.0976	22.96	7600	0.0630	0.0387
0.0942	24.17	8000	0.0631	0.0380
0.0924	25.38	8400	0.0645	0.0374
0.0862	26.59	8800	0.0677	0.0402
0.0831	27.79	9200	0.0680	0.0393
0.077	29.0	9600	0.0677	0.0380

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.1+cu102
• Datasets 1.13.3
• Tokenizers 0.10.3