fine-tune-Wav2Vec2-XLS-R-300M-Indonesia-1

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_10_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.2114
• Wer: 0.1762

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: 0.0001
• train_batch_size: 36
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 72
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 60
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9372	4.0	460	2.8359	1.0
2.2755	8.0	920	0.3708	0.3983
0.6614	12.0	1380	0.2433	0.2636
0.5071	16.0	1840	0.2347	0.2395
0.4516	20.0	2300	0.2213	0.2185
0.4206	24.0	2760	0.2222	0.2008
0.3844	28.0	3220	0.2072	0.1887
0.3678	32.0	3680	0.2071	0.1886
0.3565	36.0	4140	0.2015	0.1851
0.3388	40.0	4600	0.2137	0.1850
0.3235	44.0	5060	0.2072	0.1791
0.3173	48.0	5520	0.2095	0.1777
0.3088	52.0	5980	0.2102	0.1784
0.3	56.0	6440	0.2164	0.1772
0.2957	60.0	6900	0.2114	0.1762

Framework versions

• Transformers 4.22.2
• Pytorch 1.10.0+cu113
• Datasets 2.5.1
• Tokenizers 0.12.1