wav2vec2-child-en-tokenizer-4

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.4709
• Wer: 0.3769

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: 0.0003
• train_batch_size: 24
• eval_batch_size: 24
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 48
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0334	1.72	100	1.4709	0.3769
0.0332	3.45	200	1.4709	0.3769
0.0343	5.17	300	1.4709	0.3769
0.032	6.9	400	1.4709	0.3769
0.0332	8.62	500	1.4709	0.3769
0.0327	10.34	600	1.4709	0.3769
0.0331	12.07	700	1.4709	0.3769
0.0334	13.79	800	1.4709	0.3769
0.0319	15.52	900	1.4709	0.3769
0.0338	17.24	1000	1.4709	0.3769
0.0321	18.97	1100	1.4709	0.3769
0.0367	20.69	1200	1.4709	0.3769
0.0331	22.41	1300	1.4709	0.3769
0.0332	24.14	1400	1.4709	0.3769
0.0347	25.86	1500	1.4709	0.3769
0.0319	27.59	1600	1.4709	0.3769
0.0302	29.31	1700	1.4709	0.3769

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu113
• Datasets 1.18.3
• Tokenizers 0.10.3