wav2vec2-random

This model is a fine-tuned version of patrickvonplaten/wav2vec2-base-random on the TIMIT_ASR - NA dataset. It achieves the following results on the evaluation set:

  • Loss: 3.1593
  • Wer: 0.8364

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: 32
  • eval_batch_size: 1
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 1000
  • num_epochs: 20.0
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
2.9043 0.69 100 2.9683 1.0
2.8537 1.38 200 2.9281 0.9997
2.7803 2.07 300 2.7330 0.9999
2.6806 2.76 400 2.5792 1.0
2.4136 3.45 500 2.4327 0.9948
2.1682 4.14 600 2.3508 0.9877
2.2577 4.83 700 2.2176 0.9773
2.355 5.52 800 2.1753 0.9542
1.8588 6.21 900 2.0650 0.8851
1.6831 6.9 1000 2.0109 0.8618
1.888 7.59 1100 1.9660 0.8418
2.0066 8.28 1200 1.9847 0.8531
1.7044 8.97 1300 1.9760 0.8527
1.3168 9.66 1400 2.0708 0.8327
1.2143 10.34 1500 2.0601 0.8419
1.6189 11.03 1600 2.0960 0.8299
1.13 11.72 1700 2.2540 0.8408
0.8001 12.41 1800 2.4260 0.8306
0.7769 13.1 1900 2.4182 0.8445
1.2165 13.79 2000 2.3666 0.8284
0.8026 14.48 2100 2.7118 0.8662
0.5148 15.17 2200 2.7957 0.8526
0.4921 15.86 2300 2.8244 0.8346
0.7629 16.55 2400 2.8944 0.8370
0.5762 17.24 2500 3.0335 0.8367
0.4076 17.93 2600 3.0776 0.8358
0.3395 18.62 2700 3.1572 0.8261
0.4862 19.31 2800 3.1319 0.8414
0.5061 20.0 2900 3.1593 0.8364

Framework versions

  • Transformers 4.12.0.dev0
  • Pytorch 1.8.1
  • Datasets 1.14.1.dev0
  • Tokenizers 0.10.3
Downloads last month
13
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 patrickvonplaten/wav2vec2-random