wav2vec2-xlsr-300M-NPSC-OH

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the NBAILAB/NPSC - 16K_MP3 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1692
• Wer: 0.1663

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: 7.5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 13
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 15.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.1638	0.66	500	3.0686	1.0
2.9311	1.31	1000	2.9208	1.0
2.4175	1.97	1500	1.5009	0.9049
1.4442	2.63	2000	0.4426	0.3783
1.2624	3.28	2500	0.3193	0.2998
1.1889	3.94	3000	0.2867	0.2630
1.1315	4.6	3500	0.2566	0.2444
1.0864	5.26	4000	0.2368	0.2294
1.093	5.91	4500	0.2240	0.2151
1.0368	6.57	5000	0.2117	0.2056
1.0178	7.23	5500	0.2020	0.1954
1.0035	7.88	6000	0.2005	0.1924
0.9759	8.54	6500	0.1971	0.1863
0.9795	9.2	7000	0.1892	0.1812
0.9601	9.85	7500	0.1863	0.1795
0.9673	10.51	8000	0.1809	0.1761
0.9233	11.17	8500	0.1818	0.1755
0.9382	11.83	9000	0.1767	0.1741
0.9242	12.48	9500	0.1743	0.1703
0.9703	13.14	10000	0.1711	0.1711
0.9139	13.8	10500	0.1718	0.1672
0.9073	14.45	11000	0.1700	0.1665

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0