shivam/xls-r-300m-hindi

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: 0.8111
• Wer: 0.5177

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
4.9733	2.59	500	5.0697	1.0
3.3839	5.18	1000	3.3518	1.0
2.0596	7.77	1500	1.3992	0.7869
1.6102	10.36	2000	1.0712	0.6754
1.4587	12.95	2500	0.9280	0.6361
1.3667	15.54	3000	0.9281	0.6155
1.3042	18.13	3500	0.9037	0.5921
1.2544	20.73	4000	0.8996	0.5824
1.2274	23.32	4500	0.8934	0.5797
1.1763	25.91	5000	0.8643	0.5760
1.149	28.5	5500	0.8251	0.5544
1.1207	31.09	6000	0.8506	0.5527
1.091	33.68	6500	0.8370	0.5366
1.0613	36.27	7000	0.8345	0.5352
1.0495	38.86	7500	0.8380	0.5321
1.0345	41.45	8000	0.8285	0.5269
1.0297	44.04	8500	0.7836	0.5141
1.027	46.63	9000	0.8120	0.5180
0.9876	49.22	9500	0.8109	0.5188

Framework versions

• Transformers 4.16.0.dev0
• Pytorch 1.10.1+cu113
• Datasets 1.18.1.dev0
• Tokenizers 0.11.0