fine-tune-Wav2Vec2-XLS-R-300M-Indonesia-3

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: 0.2856
• Wer: 0.1931

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
2.8503	4.0	460	1.2215	0.8013
0.2917	8.0	920	0.2824	0.3055
0.1487	12.0	1380	0.2719	0.2593
0.0904	16.0	1840	0.2536	0.2448
0.0741	20.0	2300	0.2620	0.2325
0.0635	24.0	2760	0.2953	0.2354
0.0541	28.0	3220	0.2683	0.2184
0.048	32.0	3680	0.2707	0.2097
0.0435	36.0	4140	0.2764	0.2111
0.0377	40.0	4600	0.2874	0.2048
0.0352	44.0	5060	0.2758	0.1999
0.0304	48.0	5520	0.2808	0.1969
0.0285	52.0	5980	0.2860	0.1943
0.0258	56.0	6440	0.2867	0.1943
0.0239	60.0	6900	0.2856	0.1931

Framework versions

• Transformers 4.22.2
• Pytorch 1.10.0+cu113
• Datasets 2.5.2
• Tokenizers 0.12.1