wav2vec2-xls-r-300m-Russian-small
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: 0.3514
- Wer: 0.4838
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: 10
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
5.512 | 1.32 | 400 | 3.2207 | 1.0 |
3.1562 | 2.65 | 800 | 3.0166 | 1.0 |
1.5211 | 3.97 | 1200 | 0.7134 | 0.8275 |
0.6724 | 5.3 | 1600 | 0.4713 | 0.6402 |
0.4693 | 6.62 | 2000 | 0.3904 | 0.5668 |
0.3693 | 7.95 | 2400 | 0.3609 | 0.5121 |
0.3004 | 9.27 | 2800 | 0.3514 | 0.4838 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3
- Downloads last month
- 4
This model can be loaded on the Inference API on-demand.
Dataset used to train emre/wav2vec2-xls-r-300m-Russian-small
Evaluation results
- Test WER on Common Voice ruself-reported48.380
- Test WER on Robust Speech Event - Dev Dataself-reported58.250
- Test WER on Robust Speech Event - Test Dataself-reported56.830