whisper-dpv-finetuned-WITH-PUNCTUATION

This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6549
• Wer: 35.3154

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: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• num_epochs: 4
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3015	1.25	1000	0.5604	37.1241
0.1215	2.49	2000	0.5759	35.5189
0.0338	3.74	3000	0.6549	35.3154

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.12.1
• Datasets 2.7.1
• Tokenizers 0.13.2