wav2vec2-base-ami_multi-nithin3

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: 1.9953
  • Wer: 0.4577

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: 3e-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: 500
  • num_epochs: 30.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.7412 1.07 2500 2.9356 0.9925
2.0224 2.13 5000 2.0951 0.5730
1.9017 3.2 7500 1.8801 0.5070
1.8356 4.27 10000 2.0530 0.4778
1.8002 5.33 12500 1.9465 0.4620
1.7424 6.4 15000 1.9561 0.4529
1.7406 7.47 17500 1.9190 0.4477
1.7046 8.53 20000 1.8138 0.4402
1.6784 9.6 22500 1.8275 0.4385
1.6657 10.67 25000 1.7603 0.4307
1.6618 11.73 27500 1.7269 0.4249
1.6037 12.8 30000 1.7071 0.4272
1.639 13.87 32500 1.6559 0.4234
1.614 14.93 35000 1.7535 0.4237
1.6044 16.0 37500 1.7945 0.4200
1.5685 17.06 40000 1.7135 0.4170
1.6194 18.13 42500 1.8712 0.4161
1.566 19.2 45000 1.8720 0.4176
1.5572 20.26 47500 1.7077 0.4135
1.5715 21.33 50000 1.7538 0.4143
1.5595 22.4 52500 1.8135 0.4133
1.5465 23.46 55000 1.8119 0.4134
1.5369 24.53 57500 1.7565 0.4086
1.5392 25.6 60000 1.7323 0.4101
1.5383 26.66 62500 1.7516 0.4097
1.5266 27.73 65000 1.7961 0.4104
1.525 28.8 67500 1.7472 0.4094
1.5779 29.86 70000 1.7600 0.4096

Framework versions

  • Transformers 4.12.0.dev0
  • Pytorch 1.9.1
  • Datasets 1.12.2.dev0
  • Tokenizers 0.10.3
Downloads last month
17
Hosted inference API
Automatic Speech Recognition
or
This model can be loaded on the Inference API on-demand.