wav2vec2-xls-r-1b-dutch

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.5441
• Wer: 0.4250

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: 0.00015
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.02
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.8501	0.55	100	2.9579	1.0000
2.8419	1.09	200	2.8602	0.9999
2.5211	1.64	300	2.1828	0.9998
1.7206	2.19	400	1.2938	0.8820
1.4862	2.73	500	1.1201	0.8171
1.2637	3.28	600	1.0254	0.8137
1.0777	3.83	700	0.8863	0.6970
1.1046	4.37	800	0.8367	0.6378
0.8848	4.92	900	0.7275	0.5733
0.8134	5.46	1000	0.6990	0.5406
0.9912	6.01	1100	0.6568	0.5436
0.6002	6.56	1200	0.6744	0.5144
0.5809	7.1	1300	0.6510	0.5122
1.0383	7.65	1400	0.6050	0.4985
0.6218	8.2	1500	0.6552	0.5096
1.004	8.74	1600	0.6137	0.4794
0.8444	9.29	1700	0.5866	0.4805
0.9496	9.84	1800	0.5898	0.4726
0.7157	10.38	1900	0.5993	0.4642
0.6619	10.93	2000	0.5908	0.4706
0.4023	11.48	2100	0.6007	0.4601
0.6407	12.02	2200	0.5682	0.4583
0.3774	12.57	2300	0.5581	0.4489
0.5975	13.11	2400	0.5645	0.4490
0.5601	13.66	2500	0.5594	0.4523
0.5614	14.21	2600	0.5666	0.4421
0.3612	14.75	2700	0.5481	0.4377
0.3519	15.3	2800	0.5526	0.4374
0.325	15.85	2900	0.5244	0.4321
0.2626	16.39	3000	0.5340	0.4285
0.2425	16.94	3100	0.5495	0.4291
0.2306	17.49	3200	0.5369	0.4267
0.2374	18.03	3300	0.5379	0.4249
0.256	18.58	3400	0.5417	0.4248
0.2035	19.13	3500	0.5435	0.4251
0.1819	19.67	3600	0.5441	0.4250

Framework versions

• Transformers 4.39.3
• Pytorch 2.7.0+cu126
• Datasets 2.16.1
• Tokenizers 0.15.2