xlsr-sl

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_17_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5561
• Wer: 0.3677
• Cer: 0.0886

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: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.5901	2.2989	100	3.8082	1.0	1.0
3.0323	4.5977	200	3.0283	0.9994	0.9801
2.6749	6.8966	300	2.7497	0.9994	0.9693
0.4857	9.1954	400	0.7014	0.6455	0.1652
0.2985	11.4943	500	0.6042	0.5448	0.1409
0.2322	13.7931	600	0.5914	0.5102	0.1322
0.1573	16.0920	700	0.6055	0.5160	0.1363
0.17	18.3908	800	0.5875	0.46	0.1177
0.1093	20.6897	900	0.5322	0.4464	0.1117
0.0874	22.9885	1000	0.5529	0.4363	0.1078
0.108	25.2874	1100	0.5894	0.4397	0.1108
0.1021	27.5862	1200	0.5827	0.4182	0.1041
0.1012	29.8851	1300	0.5558	0.4113	0.1025
0.0239	32.1839	1400	0.5576	0.4013	0.1005
0.0449	34.4828	1500	0.5787	0.4138	0.1013
0.0709	36.7816	1600	0.5299	0.3799	0.0917
0.0372	39.0805	1700	0.5547	0.3827	0.0950
0.055	41.3793	1800	0.5421	0.3702	0.0892
0.0086	43.6782	1900	0.5342	0.3626	0.0881
0.0263	45.9770	2000	0.5646	0.3698	0.0899
0.0309	48.2759	2100	0.5561	0.3677	0.0886

Framework versions

• Transformers 4.42.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1