fine-tune-wav2vec2-large-xls-r-300m-sw
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_11_0 swahili dataset. It achieves the following results on the evaluation set:
- Loss: 1.2834
- Wer: 0.5834
Model description
This model is fine-tuned for general swahili speech recognition tasks. You can watch our hour long webinar and see the slides on this work.
Intended uses & limitations
The intention is to transcribe general swahili speeches. With further development, we'll fine-tune the model for domain-specific (we are focused on hospital tasks) swahili conversations.
Training and evaluation data
To appreciate the transformation we did on the data, you can read our blog on data preparation.
Training procedure
We also documented some lessons from the fine-tuning exercise.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0003
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- 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: 9
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| No log | 1.72 | 200 | 3.0092 | 1.0 |
| 4.1305 | 3.43 | 400 | 2.9159 | 1.0 |
| 4.1305 | 5.15 | 600 | 1.4301 | 0.7040 |
| 0.9217 | 6.87 | 800 | 1.3143 | 0.6529 |
| 0.9217 | 8.58 | 1000 | 1.2834 | 0.5834 |
Framework versions
- Transformers 4.27.0
- Pytorch 1.13.1+cu116
- Datasets 2.10.1
- Tokenizers 0.13.2