shivam/wav2vec2-xls-r-hindi

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

• Loss: 1.2282
• Wer: 0.6838

Evaluation results on Common Voice 7 "test" (Running ./eval.py):

With LM

• WER: 52.30
• CER: 26.09

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
5.3155	3.4	500	4.5582	1.0
3.3369	6.8	1000	3.4269	1.0
2.1785	10.2	1500	1.7191	0.8831
1.579	13.6	2000	1.3604	0.7647
1.3773	17.01	2500	1.2737	0.7519
1.3165	20.41	3000	1.2457	0.7401
1.2274	23.81	3500	1.3617	0.7301
1.1787	27.21	4000	1.2068	0.7010
1.1467	30.61	4500	1.2416	0.6946
1.0801	34.01	5000	1.2312	0.6990
1.0709	37.41	5500	1.2984	0.7138
1.0307	40.81	6000	1.2049	0.6871
1.0003	44.22	6500	1.1956	0.6841
1.004	47.62	7000	1.2101	0.6793

Framework versions

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