<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
Dansk-wav2vec2-stt
This model is a fine-tuned version of Siyam/Dansk-wav2vec21 on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 0.7500
- Wer: 0.3929
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: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- 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
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
0.0298 | 4.26 | 400 | 0.8420 | 0.4579 |
0.0479 | 8.51 | 800 | 0.8713 | 0.4461 |
0.0387 | 12.77 | 1200 | 0.8307 | 0.4404 |
0.0336 | 17.02 | 1600 | 0.8322 | 0.4144 |
0.0322 | 21.28 | 2000 | 0.7493 | 0.4081 |
0.0288 | 25.53 | 2400 | 0.7361 | 0.3951 |
0.0264 | 29.79 | 2800 | 0.7500 | 0.3929 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu113
- Datasets 2.1.0
- Tokenizers 0.10.3