wav2vec2-xls-r-300m-breton

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the mozilla-foundation/common_voice_15_0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.9939
• Wer: 0.4680
• Cer: 0.1711

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: 6e-05
• 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_ratio: 0.08
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
10.4347	2.56	250	3.7906	1.0	1.0
3.2144	5.13	500	3.0622	1.0	1.0
2.6193	7.69	750	1.6063	0.9122	0.4042
1.3658	10.26	1000	1.0917	0.7061	0.2474
0.9614	12.82	1250	0.9805	0.6384	0.2245
0.7915	15.38	1500	0.9340	0.5989	0.2124
0.6728	17.95	1750	0.9135	0.5735	0.2016
0.5889	20.51	2000	0.9381	0.5443	0.1944
0.5374	23.08	2250	0.9517	0.5372	0.1918
0.4972	25.64	2500	0.9171	0.5134	0.1828
0.4537	28.21	2750	0.9457	0.5043	0.1804
0.4217	30.77	3000	0.9479	0.4957	0.1789
0.4082	33.33	3250	0.9400	0.4906	0.1762
0.4005	35.9	3500	0.9710	0.4881	0.1770
0.3702	38.46	3750	0.9922	0.4792	0.1757
0.3473	41.03	4000	0.9724	0.4760	0.1734
0.3405	43.59	4250	0.9860	0.4726	0.1722
0.325	46.15	4500	0.9888	0.4679	0.1705
0.3193	48.72	4750	0.9939	0.4680	0.1711

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.0
• Tokenizers 0.15.0