wav2vec2-large-960h-lv60-self-with-wikipedia-lm-timit

This model is a fine-tuned version of gxbag/wav2vec2-large-960h-lv60-self-with-wikipedia-lm on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0889
• Wer: 0.4976

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: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
7.7911	2.02	250	3.0896	1.0
1.3854	4.03	500	0.0704	0.5052
0.1926	6.05	750	0.0678	0.5010
0.1472	8.06	1000	0.0794	0.5157
0.1326	10.08	1250	0.0937	0.5031
0.104	12.1	1500	0.0859	0.5055
0.0754	14.11	1750	0.0903	0.5031
0.0624	16.13	2000	0.0927	0.5034
0.0594	18.14	2250	0.0929	0.5016
0.057	20.16	2500	0.0873	0.5039
0.0476	22.18	2750	0.0974	0.5055
0.0382	24.19	3000	0.0886	0.5003
0.0329	26.21	3250	0.0832	0.4987
0.032	28.22	3500	0.0889	0.4976

Framework versions

• Transformers 4.23.0.dev0
• Pytorch 1.13.0.dev20220624+cu113
• Datasets 2.5.2.dev0
• Tokenizers 0.12.1