whisper-base-af-za-V4-Ari
This model is a fine-tuned version of openai/whisper-base on the Google FLEURS dataset. It achieves the following results on the evaluation set:
- eval_loss: 1.0084
- eval_wer: 32.0267
- eval_runtime: 152.7461
- eval_samples_per_second: 6.154
- eval_steps_per_second: 0.386
- epoch: 51.14
- step: 4500
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: 1e-05
- train_batch_size: 64
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 5000
- mixed_precision_training: Native AMP
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu116
- Datasets 2.8.1.dev0
- Tokenizers 0.13.2