english-filipino-wav2vec2-l-xls-r-test

This model is a fine-tuned version of jonatasgrosman/wav2vec2-large-xlsr-53-english on the filipino_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.5795
• Wer: 0.3996

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: 0.001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.0751	2.09	400	2.4744	0.9804
0.7852	4.19	800	0.5836	0.5620
0.3751	6.28	1200	0.4873	0.4658
0.2578	8.38	1600	0.5725	0.5289
0.1897	10.47	2000	0.5342	0.4856
0.1394	12.57	2400	0.5677	0.4761
0.1048	14.66	2800	0.5708	0.4415
0.0848	16.75	3200	0.5908	0.4374
0.0652	18.85	3600	0.5795	0.3996

Framework versions

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