Malaya-speech_fine-tune_realcase_27_Jun
This model is a fine-tuned version of malay-huggingface/wav2vec2-xls-r-300m-mixed on the uob_singlish dataset. It achieves the following results on the evaluation set:
- Loss: 0.9159
- Wer: 0.3819
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: 0.0002
- train_batch_size: 2
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 10
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 1.3176 | 1.82 | 20 | 0.8928 | 0.3542 |
| 0.6716 | 3.64 | 40 | 0.9123 | 0.3681 |
| 0.3484 | 5.45 | 60 | 0.9509 | 0.3681 |
| 0.3064 | 7.27 | 80 | 0.9227 | 0.3958 |
| 0.3017 | 9.09 | 100 | 0.9159 | 0.3819 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu113
- Datasets 1.18.3
- Tokenizers 0.10.3