english-filipino-wav2vec2-l-xls-r-test-07

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: 0.6768
• Wer: 0.3755

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: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9255	2.09	400	0.7742	0.7694
0.5792	4.19	800	0.5368	0.5250
0.3611	6.28	1200	0.4796	0.4718
0.2742	8.38	1600	0.5308	0.4764
0.201	10.47	2000	0.5885	0.4723
0.164	12.57	2400	0.5595	0.4750
0.1374	14.66	2800	0.5836	0.4366
0.1138	16.75	3200	0.6110	0.4628
0.0991	18.85	3600	0.6179	0.4174
0.0837	20.94	4000	0.6681	0.4170
0.0722	23.04	4400	0.6665	0.4103
0.0576	25.13	4800	0.7538	0.4068
0.052	27.23	5200	0.6808	0.3844
0.0449	29.32	5600	0.6768	0.3755

Framework versions

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