librispeech-100h-supervised-aug

This model is a fine-tuned version of Kuray107/librispeech-5h-supervised on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0776
• Wer: 0.0327

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: 32
• 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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3099	1.12	1000	0.0748	0.0521
0.1873	2.24	2000	0.0674	0.0440
0.146	3.36	3000	0.0671	0.0406
0.1233	4.48	4000	0.0619	0.0381
0.1098	5.61	5000	0.0618	0.0381
0.0985	6.73	6000	0.0590	0.0355
0.0907	7.85	7000	0.0659	0.0352
0.0837	8.97	8000	0.0679	0.0359
0.0762	10.09	9000	0.0701	0.0349
0.0707	11.21	10000	0.0715	0.0348
0.0666	12.33	11000	0.0719	0.0346
0.0631	13.45	12000	0.0746	0.0347
0.0593	14.57	13000	0.0757	0.0340
0.0554	15.7	14000	0.0746	0.0337
0.053	16.82	15000	0.0757	0.0331
0.0525	17.94	16000	0.0752	0.0327
0.0514	19.06	17000	0.0776	0.0327

Framework versions

• Transformers 4.16.2
• Pytorch 1.10.2
• Datasets 1.18.2
• Tokenizers 0.10.3