wav2vec2-large-xls-r-300m-dutch-baseline

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: 0.5107
• Wer: 0.2674
• Cer: 0.0863

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: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.655	1.31	400	0.9337	0.7332	0.2534
0.42	2.61	800	0.5018	0.4115	0.1374
0.2267	3.92	1200	0.4776	0.3791	0.1259
0.1624	5.23	1600	0.4807	0.3590	0.1208
0.135	6.54	2000	0.4899	0.3417	0.1121
0.1179	7.84	2400	0.5096	0.3445	0.1133
0.1035	9.15	2800	0.4563	0.3455	0.1129
0.092	10.46	3200	0.5061	0.3382	0.1127
0.0804	11.76	3600	0.4969	0.3285	0.1088
0.0748	13.07	4000	0.5274	0.3380	0.1114
0.0669	14.38	4400	0.5201	0.3115	0.1028
0.0588	15.69	4800	0.5238	0.3212	0.1054
0.0561	16.99	5200	0.5273	0.3185	0.1044
0.0513	18.3	5600	0.5577	0.3032	0.1010
0.0476	19.61	6000	0.5298	0.3050	0.1008
0.0408	20.91	6400	0.5725	0.2982	0.0984
0.0376	22.22	6800	0.5605	0.2953	0.0966
0.0339	23.53	7200	0.5419	0.2865	0.0938
0.0315	24.84	7600	0.5530	0.2782	0.0915
0.0286	26.14	8000	0.5354	0.2788	0.0917
0.0259	27.45	8400	0.5245	0.2715	0.0878
0.0231	28.76	8800	0.5107	0.2674	0.0863

Framework versions

• Transformers 4.17.0
• Pytorch 1.12.0+cu102
• Datasets 2.7.1
• Tokenizers 0.13.2