wav2vec2-large-xlsr-53-W2V2-TATAR-SMALL

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.4714
• Wer: 0.5316

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
6.2446	1.17	400	3.2621	1.0
1.739	2.35	800	0.5832	0.7688
0.4718	3.52	1200	0.4785	0.6824
0.3574	4.69	1600	0.4814	0.6792
0.2946	5.86	2000	0.4484	0.6506
0.2674	7.04	2400	0.4612	0.6225
0.2349	8.21	2800	0.4600	0.6050
0.2206	9.38	3200	0.4772	0.6048
0.2072	10.56	3600	0.4676	0.6106
0.1984	11.73	4000	0.4816	0.6079
0.1793	12.9	4400	0.4616	0.5836
0.172	14.08	4800	0.4808	0.5860
0.1624	15.25	5200	0.4854	0.5820
0.156	16.42	5600	0.4609	0.5656
0.1448	17.59	6000	0.4926	0.5817
0.1406	18.77	6400	0.4638	0.5654
0.1337	19.94	6800	0.4731	0.5652
0.1317	21.11	7200	0.4861	0.5639
0.1179	22.29	7600	0.4766	0.5521
0.1197	23.46	8000	0.4824	0.5584
0.1096	24.63	8400	0.5006	0.5559
0.1038	25.81	8800	0.4994	0.5440
0.0992	26.98	9200	0.4867	0.5405
0.0984	28.15	9600	0.4798	0.5361
0.0943	29.33	10000	0.4714	0.5316

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.14.0
• Tokenizers 0.10.3