wav2vec2-large-lv60-ami_multi-tune_dropout_0.0001_8

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: 1.4880
  • Wer: 0.4295

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: 0.0001
  • 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: 10.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.8417 0.86 1000 2.6883 0.9997
1.5626 1.72 2000 1.4253 0.4517
1.4476 2.59 3000 1.3356 0.4157
1.3874 3.45 4000 1.2814 0.4073
1.3391 4.31 5000 1.2700 0.4044
1.2983 5.17 6000 1.2423 0.3967
1.2618 6.03 7000 1.2429 0.3879
1.2414 6.9 8000 1.2290 0.3878
1.2286 7.76 9000 1.2301 0.3882
1.2254 8.62 10000 1.2140 0.3885
1.2257 9.48 11000 1.2154 0.3840

Framework versions

  • Transformers 4.12.0.dev0
  • Pytorch 1.9.1
  • Datasets 1.12.2.dev0
  • Tokenizers 0.10.3
Downloads last month
16
Hosted inference API
Automatic Speech Recognition
or
This model can be loaded on the Inference API on-demand.