Edit model card

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
Downloads last month
31
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train shivam/wav2vec2-xls-r-hindi

Evaluation results