wav2vec2-base-ami_multi-nithin6

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

  • Loss: 1.7654
  • Wer: 0.4952

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: 0.0005
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 64
  • 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
1.3222 4.31 2500 1.4875 0.5021
1.164 8.62 5000 1.4255 0.4816
1.0753 12.93 7500 1.4086 0.4717
0.9196 17.24 10000 1.4163 0.4695
0.8326 21.55 12500 1.5326 0.4650
0.7306 25.86 15000 1.5793 0.4670
0.5763 30.17 17500 1.7485 0.4728
0.4869 34.48 20000 1.9050 0.4797
0.4183 38.79 22500 2.1386 0.4835

Framework versions

  • Transformers 4.12.0.dev0
  • Pytorch 1.9.1
  • Datasets 1.12.2.dev0
  • Tokenizers 0.10.3
Downloads last month
26
Hosted inference API
Automatic Speech Recognition
or
This model can be loaded on the Inference API on-demand.