wav2vec2-large-xlsr-korean-demo-colab-2

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2481
• Wer: 0.2480

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: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• 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
4.7387	2.12	400	3.1791	1.0
1.3766	4.23	800	0.4876	0.5264
0.476	6.35	1200	0.2955	0.3648
0.3209	8.46	1600	0.2926	0.3473
0.2591	10.58	2000	0.2723	0.3094
0.2055	12.7	2400	0.2746	0.3027
0.1802	14.81	2800	0.2672	0.2976
0.1552	16.93	3200	0.2822	0.2807
0.1413	19.05	3600	0.2652	0.2856
0.1232	21.16	4000	0.2631	0.2655
0.1146	23.28	4400	0.2561	0.2574
0.1086	25.4	4800	0.2461	0.2527
0.0944	27.51	5200	0.2521	0.2535
0.0881	29.63	5600	0.2481	0.2480

Framework versions

• Transformers 4.21.0
• Pytorch 1.12.0+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1