wav2vec2-xls-r-300m-ab-CV8
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 0.2105
- Wer: 0.5474
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: 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: 300
- num_epochs: 15
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
4.7729 | 0.63 | 500 | 3.0624 | 1.0021 |
2.7348 | 1.26 | 1000 | 1.0460 | 0.9815 |
1.2756 | 1.9 | 1500 | 0.4618 | 0.8309 |
1.0419 | 2.53 | 2000 | 0.3725 | 0.7449 |
0.9491 | 3.16 | 2500 | 0.3368 | 0.7345 |
0.9006 | 3.79 | 3000 | 0.3014 | 0.6936 |
0.8519 | 4.42 | 3500 | 0.2852 | 0.6767 |
0.8243 | 5.06 | 4000 | 0.2701 | 0.6504 |
0.7902 | 5.69 | 4500 | 0.2641 | 0.6221 |
0.7767 | 6.32 | 5000 | 0.2549 | 0.6192 |
0.7516 | 6.95 | 5500 | 0.2515 | 0.6179 |
0.737 | 7.59 | 6000 | 0.2408 | 0.5963 |
0.7217 | 8.22 | 6500 | 0.2429 | 0.6261 |
0.7101 | 8.85 | 7000 | 0.2366 | 0.5687 |
0.6922 | 9.48 | 7500 | 0.2277 | 0.5680 |
0.6866 | 10.11 | 8000 | 0.2242 | 0.5847 |
0.6703 | 10.75 | 8500 | 0.2222 | 0.5803 |
0.6649 | 11.38 | 9000 | 0.2247 | 0.5765 |
0.6513 | 12.01 | 9500 | 0.2182 | 0.5644 |
0.6369 | 12.64 | 10000 | 0.2128 | 0.5508 |
0.6425 | 13.27 | 10500 | 0.2132 | 0.5514 |
0.6399 | 13.91 | 11000 | 0.2116 | 0.5495 |
0.6208 | 14.54 | 11500 | 0.2105 | 0.5474 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.18.1
- Tokenizers 0.10.3
- Downloads last month
- 16
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.