Shiry/whisper-large-v2-he-1

Whisper Large V2 Hebrew

This model is a fine-tuned version of openai/whisper-large-v2 on the google/fleurs he_il dataset. It achieves the following results on the evaluation set:

• Loss: 0.4106
• Wer: 27.2504

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-06
• train_batch_size: 128
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 512
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 10
• training_steps: 200
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.425	24.01	50	0.4106	27.2504
0.1906	49.01	100	0.4420	29.0131
0.0982	74.01	150	0.4795	30.3063
0.0717	99.01	200	0.4945	30.8915

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1+cu117
• Datasets 2.8.1.dev0
• Tokenizers 0.13.2