wav2vec2-large-xlsr-korean-demo-test2

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: 1.0566
• Wer: 0.5224

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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
31.2541	0.3	400	5.4002	1.0
4.9419	0.59	800	5.3336	1.0
4.8926	0.89	1200	5.0531	1.0
4.7218	1.19	1600	4.5172	1.0
4.0218	1.49	2000	3.1418	0.9518
3.0654	1.78	2400	2.4376	0.9041
2.6226	2.08	2800	2.0151	0.8643
2.2944	2.38	3200	1.8025	0.8290
2.1872	2.67	3600	1.6469	0.7962
2.0747	2.97	4000	1.5165	0.7714
1.8479	3.27	4400	1.4281	0.7694
1.8288	3.57	4800	1.3791	0.7326
1.801	3.86	5200	1.3328	0.7177
1.6723	4.16	5600	1.2954	0.7192
1.5925	4.46	6000	1.3137	0.6953
1.5709	4.75	6400	1.2086	0.6973
1.5294	5.05	6800	1.1811	0.6730
1.3844	5.35	7200	1.2053	0.6769
1.3906	5.65	7600	1.1287	0.6556
1.4088	5.94	8000	1.1251	0.6466
1.2989	6.24	8400	1.1577	0.6546
1.2523	6.54	8800	1.0643	0.6377
1.2651	6.84	9200	1.0865	0.6417
1.2209	7.13	9600	1.0981	0.6272
1.1435	7.43	10000	1.1195	0.6317
1.1616	7.73	10400	1.0672	0.6327
1.1272	8.02	10800	1.0413	0.6248
1.043	8.32	11200	1.0555	0.6233
1.0523	8.62	11600	1.0372	0.6178
1.0208	8.92	12000	1.0170	0.6128
0.9895	9.21	12400	1.0354	0.5934
0.95	9.51	12800	1.1019	0.6039
0.9705	9.81	13200	1.0229	0.5855
0.9202	10.1	13600	1.0364	0.5919
0.8644	10.4	14000	1.0721	0.5984
0.8641	10.7	14400	1.0383	0.5905
0.8924	11.0	14800	0.9947	0.5760
0.7914	11.29	15200	1.0270	0.5885
0.7882	11.59	15600	1.0271	0.5741
0.8116	11.89	16000	0.9937	0.5741
0.7584	12.18	16400	0.9924	0.5626
0.7051	12.48	16800	1.0023	0.5572
0.7232	12.78	17200	1.0479	0.5512
0.7149	13.08	17600	1.0475	0.5765
0.6579	13.37	18000	1.0218	0.5552
0.6615	13.67	18400	1.0339	0.5631
0.6629	13.97	18800	1.0239	0.5621
0.6221	14.26	19200	1.0331	0.5537
0.6159	14.56	19600	1.0640	0.5532
0.6032	14.86	20000	1.0192	0.5567
0.5748	15.16	20400	1.0093	0.5507
0.5614	15.45	20800	1.0458	0.5472
0.5626	15.75	21200	1.0318	0.5398
0.5429	16.05	21600	1.0112	0.5278
0.5407	16.34	22000	1.0120	0.5278
0.511	16.64	22400	1.0335	0.5249
0.5316	16.94	22800	1.0146	0.5348
0.4949	17.24	23200	1.0287	0.5388
0.496	17.53	23600	1.0229	0.5348
0.4986	17.83	24000	1.0094	0.5313
0.4787	18.13	24400	1.0620	0.5234
0.4508	18.42	24800	1.0401	0.5323
0.4754	18.72	25200	1.0543	0.5303
0.4584	19.02	25600	1.0433	0.5194
0.4431	19.32	26000	1.0597	0.5249
0.4448	19.61	26400	1.0548	0.5229
0.4475	19.91	26800	1.0566	0.5224

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1