wav2vec2-large-xls-r-300m-Swedish

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

• Loss: 0.3641
• Wer: 0.2473
• Cer: 0.0758

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7.5e-05
• train_batch_size: 64
• eval_batch_size: 8
• 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: 1000
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
6.1097	5.49	500	3.1422	1.0	1.0
2.985	10.98	1000	1.7357	0.9876	0.4125
1.0363	16.48	1500	0.4773	0.3510	0.1047
0.6111	21.97	2000	0.3937	0.2998	0.0910
0.4942	27.47	2500	0.3779	0.2776	0.0844
0.4421	32.96	3000	0.3745	0.2630	0.0807
0.4018	38.46	3500	0.3685	0.2553	0.0781
0.3759	43.95	4000	0.3618	0.2488	0.0761
0.3646	49.45	4500	0.3641	0.2473	0.0758

Framework versions

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