arampacha/wav2vec2-xls-r-300m-hy-cv

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - HY-AM dataset. It achieves the following results on the evaluation set:

• Loss: 0.5891
• Wer: 0.6569

Note: If you aim for best performance use this model. It is trained using noizy student procedure and achieves considerably better results.

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: 0.0001
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• training_steps: 1200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
9.167	16.67	100	3.5599	1.0
3.2645	33.33	200	3.1771	1.0
3.1509	50.0	300	3.1321	1.0
3.0757	66.67	400	2.8594	1.0
2.5274	83.33	500	1.5286	0.9797
1.6826	100.0	600	0.8058	0.7974
1.2868	116.67	700	0.6713	0.7279
1.1262	133.33	800	0.6308	0.7034
1.0408	150.0	900	0.6056	0.6745
0.9617	166.67	1000	0.5891	0.6569
0.9196	183.33	1100	0.5913	0.6432
0.8853	200.0	1200	0.5924	0.6347

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0