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