Whisper base arabic
It achieves the following results on the evaluation set:
- Loss: 0.44
- Wer: 34.7
Training and evaluation data
Train set:
- mozilla-foundation/common_voice_16_0 ar [train+validation]
- BelalElhossany/mgb2_audios_transcriptions_non_overlap
- nadsoft/Jordan-Audio
cross validation set: 600 samples in total from the 3 sets to save time during training as colab free tier was used to train the model. note: evaluate accuracy in the way you see fit.
Training procedure
removed arabic (حركات) from the texts. trained the model on the combined dataset for 6 epochs, the best one being the fifth so the model is basically the 5th epoch.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 1
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.4603 | 1 | 1437 | 0.4931 | 45.8857 |
| 0.2867 | 2 | 2874 | 0.4493 | 36.9973 |
| 0.2494 | 3 | 4311 | 0.4219 | 43.5553 |
| 0.1435 | 4 | 5748 | 0.4408 | 40.2351 |
| 0.1345 | 5 | 7185 | 0.4407 | 34.7081 |