w2v2-libri

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

• Loss: 1.7315
• Wer: 0.5574

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=(0.9,0.98) and epsilon=1e-07
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 3000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
7.1828	50.0	200	3.0563	1.0
2.8849	100.0	400	2.9023	1.0
1.5108	150.0	600	1.1468	0.6667
0.1372	200.0	800	1.3749	0.6279
0.0816	250.0	1000	1.3985	0.6224
0.0746	300.0	1200	1.5285	0.6141
0.0556	350.0	1400	1.5496	0.5920
0.0644	400.0	1600	1.6263	0.5947
0.0546	450.0	1800	1.6803	0.5906
0.0491	500.0	2000	1.6155	0.5837
0.0518	550.0	2200	1.6784	0.5698
0.0314	600.0	2400	1.6050	0.5602
0.0048	650.0	2600	1.7703	0.5546
0.0042	700.0	2800	1.7135	0.5615
0.0025	750.0	3000	1.7315	0.5574

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.0+cu116
• Datasets 1.18.3
• Tokenizers 0.13.2