wav2vec2-bloom-speech-bzi
Model description
- Homepage: SIL AI
- Point of Contact: SIL AI email
- Source Data: Bloom Library
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the SIL-AI/BLOOM-SPEECH - BZI (Bisu) dataset. It achieves the following results on the evaluation set:
- Loss: 0.1160
- Wer: 0.0802
- Cer: 0.0159
Users should refer to the original model for tutorials on using a trained model for inference.
Intended uses & limitations
Users of this model must abide by the SIL RAIL-M License.
This model is created as a proof of concept and no guarantees are made regarding the performance of the model is specific situations.
Training and evaluation data
Training, Validation, and Test datasets were generated from the same corpus, ensuring that no duplicate files were used.
Training procedure
Standard finetuning of XLS-R was used based on the examples in the Hugging Face Transformers Github
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: 250
- num_epochs: 1000.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
|---|---|---|---|---|---|
| No log | 5.81 | 250 | 1.1561 | 0.7510 | 0.2887 |
| 2.7319 | 11.63 | 500 | 0.1290 | 0.1193 | 0.0246 |
| 2.7319 | 17.44 | 750 | 0.1441 | 0.1420 | 0.0306 |
| 0.1059 | 23.26 | 1000 | 0.1160 | 0.0802 | 0.0159 |
| 0.1059 | 29.07 | 1250 | 0.1252 | 0.1008 | 0.0177 |
| 0.0635 | 34.88 | 1500 | 0.1105 | 0.0905 | 0.0185 |
| 0.0635 | 40.7 | 1750 | 0.1699 | 0.0967 | 0.0185 |
Framework versions
- Transformers 4.21.0.dev0
- Pytorch 1.9.0+cu111
- Datasets 2.2.2
- Tokenizers 0.12.1
