anuragshas/wav2vec2-xls-r-300m-ur-cv9-with-lm

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - UR dataset. It achieves the following results on the evaluation set:

• Loss: 0.4147
• Wer: 0.3172
• Cer: 0.1050

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: 7.5e-05
• 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: 5108
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
3.2894	7.83	400	3.1501	1.0	1.0
1.8586	15.68	800	0.8871	0.6721	0.2402
1.3431	23.52	1200	0.5813	0.5502	0.1939
1.2052	31.37	1600	0.4956	0.4788	0.1665
1.1097	39.21	2000	0.4447	0.4143	0.1397
1.0528	47.06	2400	0.4439	0.3961	0.1333
0.9939	54.89	2800	0.4348	0.4014	0.1379
0.9441	62.74	3200	0.4236	0.3653	0.1223
0.913	70.58	3600	0.4309	0.3475	0.1157
0.8678	78.43	4000	0.4270	0.3337	0.1110
0.8414	86.27	4400	0.4158	0.3220	0.1070
0.817	94.12	4800	0.4185	0.3231	0.1072

Framework versions

• Transformers 4.19.0.dev0
• Pytorch 1.11.0+cu102
• Datasets 2.1.1.dev0
• Tokenizers 0.12.1