xls-r-300m-hi

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

• Loss: 0.7522
• Wer: 1.0091

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2000
• num_epochs: 50.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.0417	2.59	500	5.1484	1.0
3.3722	5.18	1000	3.3380	1.0001
1.9752	7.77	1500	1.3910	1.0074
1.5868	10.36	2000	1.0298	1.0084
1.4413	12.95	2500	0.9313	1.0175
1.3296	15.54	3000	0.8966	1.0194
1.2746	18.13	3500	0.8875	1.0097
1.2147	20.73	4000	0.8746	1.0089
1.1774	23.32	4500	0.8383	1.0198
1.129	25.91	5000	0.7848	1.0167
1.0995	28.5	5500	0.7992	1.0210
1.0665	31.09	6000	0.7878	1.0107
1.0321	33.68	6500	0.7653	1.0082
1.0068	36.27	7000	0.7635	1.0065
0.9916	38.86	7500	0.7728	1.0090
0.9735	41.45	8000	0.7688	1.0070
0.9745	44.04	8500	0.7455	1.0097
0.9677	46.63	9000	0.7605	1.0099
0.9313	49.22	9500	0.7527	1.0097

Framework versions

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