# CaptainA_XLS-R_entropy-10_v0
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the Finnish Parliament Corpus with 210h of slow and clean samples. It achieves the following results on the evaluation set (70h of the slow and clean samples from the same corpus):
- Loss: 196.9006
- Cer: 0.0178
- Wer: 0.0592
Model description
This model is used in the CaptainA app.
Intended uses & limitations
The model was fine-tuned with entropy regularization to improve generalization for the Finnish L2 MDD task. Even though the model was trained for the purpose of MDD for L2 Finnish speakers, it was not fine-tuned with any L2 data due to the lack of proper corpus for Finnish L2 MDD.
Therefore, it is important to note that this model is NOT intended for Finnish L2 ASR and will need to be improved further even for the MDD task (hence the model version is 0).
Because the model was fine-tuned with the Finnish Parliament Corpus, it has the same biases from that corpus. The biases are more notable since the model's intended uses are for L2 Finnish speakers. More detail can be read in the Master's thesis: A Mobile App For Practicing Finnish Pronunciation Using Wav2vec 2.0
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 128
- eval_batch_size: 128
- seed: 1011
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 10.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Cer | Validation Loss | Wer |
|---|---|---|---|---|---|
| 409.2657 | 0.47 | 500 | 1.0 | 409.9399 | 1.0 |
| 242.0796 | 0.94 | 1000 | 0.0762 | 222.4043 | 0.4204 |
| 214.4323 | 1.4 | 1500 | 0.0342 | 205.3006 | 0.1620 |
| 208.1767 | 1.87 | 2000 | 0.0261 | 201.7304 | 0.1095 |
| 205.3693 | 2.34 | 2500 | 0.0248 | 200.3012 | 0.1037 |
| 204.3477 | 2.81 | 3000 | 0.0219 | 199.3383 | 0.0830 |
| 202.9748 | 3.28 | 3500 | 0.0207 | 199.1589 | 0.0782 |
| 201.9818 | 3.75 | 4000 | 0.0207 | 198.5560 | 0.0769 |
| 201.8992 | 4.21 | 4500 | 0.0201 | 198.0990 | 0.0724 |
| 201.6079 | 4.68 | 5000 | 0.0209 | 197.8516 | 0.0712 |
| 200.6187 | 5.15 | 5500 | 0.0191 | 197.6185 | 0.0667 |
| 200.5608 | 5.62 | 6000 | 0.0189 | 197.5194 | 0.0658 |
| 200.1649 | 6.09 | 6500 | 0.0191 | 197.3655 | 0.0641 |
| 200.1713 | 6.55 | 7000 | 0.0186 | 197.2977 | 0.0629 |
| 200.1245 | 7.02 | 7500 | 0.0193 | 197.0914 | 0.0638 |
| 199.5289 | 7.49 | 8000 | 0.0181 | 197.0704 | 0.0608 |
| 199.4458 | 7.96 | 8500 | 0.0183 | 196.9986 | 0.0606 |
| 199.1502 | 8.43 | 9000 | 0.0178 | 197.0260 | 0.0590 |
| 199.4437 | 8.9 | 9500 | 0.0180 | 196.9412 | 0.0595 |
| 198.8669 | 9.36 | 10000 | 0.0180 | 196.8834 | 0.0600 |
| 199.1329 | 9.83 | 10500 | 0.0178 | 196.9176 | 0.0591 |
Framework versions
- Transformers 4.18.0
- Pytorch 1.12.0.dev20220305
- Datasets 1.18.4.dev0
- Tokenizers 0.11.6
- Downloads last month
- 8