<!-- 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. -->
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HI dataset. It achieves the following results on the evaluation set:
- Loss: 0.7805
- Wer: 0.4340
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
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 8000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| No log | 1.36 | 400 | 1.9130 | 0.9244 |
| 5.0013 | 2.71 | 800 | 0.7789 | 0.5944 |
| 0.6544 | 4.07 | 1200 | 0.7298 | 0.5852 |
| 0.4021 | 5.42 | 1600 | 0.6978 | 0.5667 |
| 0.3003 | 6.78 | 2000 | 0.6764 | 0.5382 |
| 0.3003 | 8.14 | 2400 | 0.7249 | 0.5463 |
| 0.2345 | 9.49 | 2800 | 0.7280 | 0.5124 |
| 0.1993 | 10.85 | 3200 | 0.7289 | 0.4690 |
| 0.1617 | 12.2 | 3600 | 0.7431 | 0.4733 |
| 0.1432 | 13.56 | 4000 | 0.7448 | 0.4733 |
| 0.1432 | 14.92 | 4400 | 0.7746 | 0.4485 |
| 0.1172 | 16.27 | 4800 | 0.7589 | 0.4742 |
| 0.1035 | 17.63 | 5200 | 0.7539 | 0.4353 |
| 0.0956 | 18.98 | 5600 | 0.7648 | 0.4495 |
| 0.0845 | 20.34 | 6000 | 0.7877 | 0.4719 |
| 0.0845 | 21.69 | 6400 | 0.7884 | 0.4434 |
| 0.0761 | 23.05 | 6800 | 0.7796 | 0.4386 |
| 0.0634 | 24.41 | 7200 | 0.7729 | 0.4306 |
| 0.0571 | 25.76 | 7600 | 0.7826 | 0.4298 |
| 0.0508 | 27.12 | 8000 | 0.7805 | 0.4340 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.1+cu102
- Datasets 1.18.3
- Tokenizers 0.11.0
