wav2vec2-large-xlsr-korean-demo-no-LM

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

• Loss: 0.3215
• Wer: 0.2209

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• 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
4.7322	1.08	400	3.1660	1.0
1.6742	2.16	800	0.5714	0.5945
0.6009	3.23	1200	0.3934	0.4298
0.4335	4.31	1600	0.3855	0.4100
0.3615	5.39	2000	0.3226	0.3525
0.2975	6.47	2400	0.3079	0.3176
0.2822	7.55	2800	0.3226	0.3091
0.2468	8.63	3200	0.2935	0.2907
0.2307	9.7	3600	0.2826	0.2728
0.2035	10.78	4000	0.2876	0.2728
0.1959	11.86	4400	0.2988	0.2667
0.1714	12.94	4800	0.3176	0.2751
0.1728	14.02	5200	0.2889	0.2649
0.1552	15.09	5600	0.2893	0.2490
0.144	16.17	6000	0.2909	0.2548
0.1402	17.25	6400	0.2999	0.2494
0.1297	18.33	6800	0.3704	0.2584
0.1268	19.41	7200	0.3464	0.2497
0.1162	20.49	7600	0.3620	0.2461
0.1117	21.56	8000	0.2935	0.2387
0.1081	22.64	8400	0.3588	0.2427
0.0984	23.72	8800	0.4317	0.2507
0.0996	24.8	9200	0.3023	0.2277
0.0925	25.88	9600	0.3224	0.2292
0.0923	26.95	10000	0.3009	0.2243
0.0839	28.03	10400	0.3118	0.2219
0.0814	29.11	10800	0.3215	0.2209

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.0+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1