wav2vec2-large-xls-r-300m-sakha

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - SAH dataset. It achieves the following results on the evaluation set:

• Loss: 0.4995
• Wer: 0.4421

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: 32
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 100.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.8597	8.47	500	0.7731	0.7211
1.2508	16.95	1000	0.5368	0.5989
1.1066	25.42	1500	0.5034	0.5533
1.0064	33.9	2000	0.4686	0.5114
0.9324	42.37	2500	0.4927	0.5056
0.876	50.85	3000	0.4734	0.4795
0.8082	59.32	3500	0.4748	0.4799
0.7604	67.8	4000	0.4949	0.4691
0.7241	76.27	4500	0.5090	0.4627
0.6739	84.75	5000	0.4967	0.4452
0.6447	93.22	5500	0.5071	0.4437

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu102
• Datasets 1.17.1.dev0
• Tokenizers 0.11.0