wav2vec2-large-lv60-ami_multi-nithin8

This model is a fine-tuned version of facebook/wav2vec2-large-lv60 on the AMI-IHM dataset. It achieves the following results on the evaluation set:

  • Loss: 1.4945
  • Wer: 0.4291

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: 0.0003
  • train_batch_size: 4
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 8
  • 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: 500
  • num_epochs: 40.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
1.336 2.16 2500 1.2807 0.4097
1.216 4.31 5000 1.2406 0.3931
1.1353 6.47 7500 1.2145 0.3801
1.0674 8.62 10000 1.1930 0.3825
1.0223 10.78 12500 1.2283 0.3907
1.009 12.93 15000 1.2266 0.3810
0.8998 15.09 17500 1.2719 0.3839
0.8912 17.24 20000 1.2889 0.3867
0.8459 19.4 22500 1.3031 0.3941
0.8193 21.55 25000 1.3543 0.3862
0.8048 23.71 27500 1.3533 0.3858
0.7663 25.86 30000 1.3941 0.3993
0.7311 28.02 32500 1.4745 0.3937
0.716 30.17 35000 1.4788 0.3989
0.6868 32.33 37500 1.4966 0.3925
0.6558 34.48 40000 1.5457 0.3901
0.6473 36.64 42500 1.5662 0.3944
0.631 38.79 45000 1.5689 0.3956

Framework versions

  • Transformers 4.12.0.dev0
  • Pytorch 1.9.1
  • Datasets 1.12.2.dev0
  • Tokenizers 0.10.3
Downloads last month
14
Hosted inference API
Automatic Speech Recognition
or
This model can be loaded on the Inference API on-demand.