wsj0-5percent-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: 0.3883
• Wer: 0.1555

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: 12
• 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: 300
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.0248	16.67	500	2.9406	1.0
2.0466	33.33	1000	0.3935	0.3300
0.1486	50.0	1500	0.3091	0.1931
0.052	66.67	2000	0.3562	0.2052
0.0309	83.33	2500	0.3252	0.1773
0.0228	100.0	3000	0.3360	0.1652
0.0177	116.67	3500	0.3423	0.1603
0.0142	133.33	4000	0.3416	0.1611
0.0119	150.0	4500	0.3663	0.1583
0.0094	166.67	5000	0.3617	0.1567
0.0093	183.33	5500	0.3738	0.1668
0.0079	200.0	6000	0.3881	0.1652
0.0065	216.67	6500	0.3752	0.1611
0.0056	233.33	7000	0.3798	0.1603
0.0057	250.0	7500	0.3944	0.1624
0.0047	266.67	8000	0.4038	0.1583
0.0041	283.33	8500	0.3928	0.1547
0.0036	300.0	9000	0.3883	0.1555

Framework versions

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