shivam/xls-r-300m-marathi

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - MR dataset. It achieves the following results on the mozilla-foundation/common_voice_8_0 mr test set:

• Without LM

  • WER: 48.53
  • CER: 10.63

• With LM

  • WER: 38.27
  • CER: 8.91

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: 400.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
4.2706	22.73	500	4.0174	1.0
3.2492	45.45	1000	3.2309	0.9908
1.9709	68.18	1500	1.0651	0.8440
1.4088	90.91	2000	0.5765	0.6550
1.1326	113.64	2500	0.4842	0.5760
0.9709	136.36	3000	0.4785	0.6013
0.8433	159.09	3500	0.5048	0.5419
0.7404	181.82	4000	0.5052	0.5339
0.6589	204.55	4500	0.5237	0.5897
0.5831	227.27	5000	0.5166	0.5447
0.5375	250.0	5500	0.5292	0.5487
0.4784	272.73	6000	0.5480	0.5596
0.4421	295.45	6500	0.5682	0.5467
0.4047	318.18	7000	0.5681	0.5447
0.3779	340.91	7500	0.5783	0.5347
0.3525	363.64	8000	0.5856	0.5367
0.3393	386.36	8500	0.5960	0.5359

Framework versions

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