Malaya-speech_fine-tune_MrBrown_20_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: 0.8868
- Wer: 0.3244
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
- lr_scheduler_warmup_steps: 500
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.8027 | 3.85 | 200 | 0.4800 | 0.2852 |
| 0.3773 | 7.69 | 400 | 0.6292 | 0.3316 |
| 0.3394 | 11.54 | 600 | 0.7376 | 0.3494 |
| 0.2653 | 15.38 | 800 | 0.9595 | 0.3137 |
| 0.1785 | 19.23 | 1000 | 0.7381 | 0.3440 |
| 0.1669 | 23.08 | 1200 | 0.9534 | 0.3529 |
| 0.0971 | 26.92 | 1400 | 0.8868 | 0.3244 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu113
- Datasets 1.18.3
- Tokenizers 0.10.3