librispeech-5h-supervised

This model is a fine-tuned version of facebook/wav2vec2-large-lv60 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2041
• Wer: 0.0624

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: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.7758	11.11	1000	0.3120	0.2337
0.1238	22.22	2000	0.1651	0.0826
0.0383	33.33	3000	0.1667	0.0712
0.023	44.44	4000	0.1893	0.0685
0.0166	55.56	5000	0.2008	0.0666
0.0131	66.67	6000	0.1942	0.0639
0.0106	77.78	7000	0.1979	0.0628
0.0091	88.89	8000	0.2027	0.0628
0.008	100.0	9000	0.2041	0.0624

Framework versions

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