wav2vec2-large-xls-r-300m-basaa

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

• Loss: 0.5975
• Wer: 0.4981

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: 7e-05
• train_batch_size: 32
• eval_batch_size: 32
• 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: 200.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.9287	15.62	500	2.8774	1.0
1.1182	31.25	1000	0.6248	0.7131
0.8329	46.88	1500	0.5573	0.5792
0.7109	62.5	2000	0.5420	0.5683
0.6295	78.12	2500	0.5166	0.5395
0.5715	93.75	3000	0.5487	0.5629
0.5016	109.38	3500	0.5370	0.5471
0.4661	125.0	4000	0.5621	0.5395
0.423	140.62	4500	0.5658	0.5248
0.3793	156.25	5000	0.5921	0.4981
0.3651	171.88	5500	0.5987	0.4888
0.3351	187.5	6000	0.6017	0.4948

Framework versions

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