diallomama/wav2vec2-xls-r-300m-ar

wav2vec2-xls-r-300m-en-ar-fr-es

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: 0.8565
• Wer: 0.4869

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: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.3938	0.59	400	3.3703	1.0
2.353	1.18	800	0.9696	0.7809
0.9859	1.77	1200	0.7031	0.6515
0.7685	2.35	1600	0.6575	0.6321
0.6892	2.94	2000	0.6030	0.5927
0.5866	3.53	2400	0.5552	0.5541
0.5496	4.12	2800	0.5805	0.5503
0.4897	4.71	3200	0.5526	0.5335
0.4671	5.3	3600	0.5622	0.5507
0.4346	5.89	4000	0.5641	0.5312
0.3859	6.48	4400	0.5685	0.5071
0.3728	7.06	4800	0.6106	0.5157
0.3243	7.65	5200	0.6782	0.5270
0.3073	8.24	5600	0.6121	0.5232
0.2748	8.83	6000	0.6318	0.5209
0.25	9.42	6400	0.6334	0.4906
0.2477	10.01	6800	0.6403	0.5169
0.2125	10.6	7200	0.6498	0.5080
0.1997	11.18	7600	0.7029	0.5153
0.1803	11.77	8000	0.6796	0.5193
0.1644	12.36	8400	0.7320	0.5080
0.1609	12.95	8800	0.6705	0.5081
0.1419	13.54	9200	0.7108	0.5120
0.1375	14.13	9600	0.7570	0.4909
0.1265	14.72	10000	0.7681	0.5044
0.1152	15.31	10400	0.8180	0.5011
0.1094	15.89	10800	0.7753	0.4947
0.0998	16.48	11200	0.8077	0.4972
0.1019	17.07	11600	0.8189	0.4921
0.0882	17.66	12000	0.8351	0.4922
0.0855	18.25	12400	0.8688	0.4902
0.0826	18.84	12800	0.8476	0.4916
0.0769	19.43	13200	0.8565	0.4869

Framework versions

• Transformers 4.28.0
• Pytorch 2.0.1+cu118
• Datasets 1.18.3
• Tokenizers 0.13.3