w2v2-xlsr-indonesian

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

• Loss: 0.2496
• Wer: 0.2116

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: 64
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 300
• num_epochs: 60.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.828	3.23	100	3.3112	1.0
2.99	6.45	200	2.8944	1.0
2.8597	9.68	300	2.8430	1.0
2.6267	12.9	400	1.6697	0.9975
1.3279	16.13	500	0.5147	0.5183
0.8697	19.35	600	0.3648	0.3825
0.7173	22.58	700	0.3172	0.3332
0.6268	25.81	800	0.3027	0.2900
0.5733	29.03	900	0.2921	0.2710
0.5283	32.26	1000	0.2731	0.2546
0.4865	35.48	1100	0.2690	0.2397
0.4559	38.71	1200	0.2581	0.2330
0.4352	41.94	1300	0.2620	0.2254
0.4265	45.16	1400	0.2496	0.2228
0.3987	48.39	1500	0.2571	0.2219
0.3977	51.61	1600	0.2514	0.2138
0.3786	54.84	1700	0.2533	0.2142
0.3763	58.06	1800	0.2496	0.2116

Framework versions

• Transformers 4.24.0
• Pytorch 1.10.0+cu113
• Datasets 1.18.3
• Tokenizers 0.13.2