Scrya/whisper-medium-vi-cv

Whisper Medium VI - CV - Augmented

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.6613
• Wer: 18.0303
• Cer: 8.3095

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: 32
• 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

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.0053	11.49	1000	0.5429	18.1290	8.4643
0.0021	22.99	2000	0.5916	18.8857	8.6538
0.0001	34.48	3000	0.6348	18.3374	8.4296
0.0001	45.98	4000	0.6508	17.9754	8.3149
0.0001	57.47	5000	0.6613	18.0303	8.3095

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.7.1.dev0
• Tokenizers 0.13.2