<!-- 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. -->
finetune_indian_asr
This model is a fine-tuned version of Harveenchadha/vakyansh-wav2vec2-indian-english-enm-700 on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.4215
- Wer: 0.3403
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: 8
- 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: 30
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 5.0566 | 3.45 | 500 | 2.9944 | 1.0 |
| 2.7241 | 6.9 | 1000 | 1.4455 | 0.7654 |
| 0.9755 | 10.34 | 1500 | 0.4299 | 0.4034 |
| 0.4624 | 13.79 | 2000 | 0.3628 | 0.3297 |
| 0.3158 | 17.24 | 2500 | 0.3835 | 0.2952 |
| 0.2604 | 20.69 | 3000 | 0.3802 | 0.2877 |
| 0.2 | 24.14 | 3500 | 0.3842 | 0.2799 |
| 1.7441 | 27.59 | 4000 | 0.4215 | 0.3403 |
Framework versions
- Transformers 4.17.0
- Pytorch 1.10.0+cu111
- Datasets 2.0.0
- Tokenizers 0.11.6
