librispeech-semi-supervised-without-LM

This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1837
• Wer: 0.0580

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 15
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0565	0.56	1000	0.1354	0.0641
0.0548	1.12	2000	0.1320	0.0628
0.0478	1.68	3000	0.1247	0.0612
0.0451	2.24	4000	0.1256	0.0613
0.0401	2.8	5000	0.1269	0.0606
0.035	3.36	6000	0.1370	0.0595
0.0344	3.92	7000	0.1280	0.0589
0.031	4.48	8000	0.1350	0.0589
0.031	5.04	9000	0.1418	0.0614
0.0278	5.61	10000	0.1382	0.0604
0.0272	6.17	11000	0.1502	0.0615
0.0246	6.73	12000	0.1443	0.0609
0.0233	7.29	13000	0.1548	0.0589
0.0224	7.85	14000	0.1547	0.0599
0.0202	8.41	15000	0.1570	0.0590
0.0199	8.97	16000	0.1564	0.0594
0.0186	9.53	17000	0.1598	0.0595
0.0187	10.09	18000	0.1657	0.0585
0.017	10.65	19000	0.1690	0.0584
0.016	11.21	20000	0.1689	0.0588
0.0156	11.77	21000	0.1745	0.0585
0.0151	12.33	22000	0.1777	0.0583
0.0144	12.89	23000	0.1778	0.0590
0.0142	13.45	24000	0.1803	0.0585
0.0137	14.01	25000	0.1796	0.0581
0.0132	14.57	26000	0.1837	0.0580

Framework versions

• Transformers 4.14.1
• Pytorch 1.10.2
• Datasets 1.18.2
• Tokenizers 0.10.3