wav2vec2-xls-r-300m-CV8-ro

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - RO dataset. It achieves the following results on the evaluation set:

• Loss: 0.1578
• Wer: 0.6040
• Cer: 0.0475

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
2.9736	3.62	500	2.9508	1.0	1.0
1.3293	7.25	1000	0.3330	0.8407	0.0862
0.956	10.87	1500	0.2042	0.6872	0.0602
0.9509	14.49	2000	0.2184	0.7088	0.0652
0.9272	18.12	2500	0.2312	0.7211	0.0703
0.8561	21.74	3000	0.2158	0.6838	0.0631
0.8258	25.36	3500	0.1970	0.6844	0.0601
0.7993	28.98	4000	0.1895	0.6698	0.0577
0.7525	32.61	4500	0.1845	0.6453	0.0550
0.7211	36.23	5000	0.1781	0.6274	0.0531
0.677	39.85	5500	0.1732	0.6188	0.0514
0.6517	43.48	6000	0.1691	0.6177	0.0503
0.6326	47.1	6500	0.1619	0.6045	0.0479

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0