multilingual-whisper-v3

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4028
• Wer: 0.2305
• Cer: 0.0739
• Bertscore Precision: 0.9438
• Bertscore Recall: 0.9421
• Bertscore F1: 0.9429

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: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use adamw_torch_fused with betas=(0.9,0.98) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 2000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Bertscore Precision	Bertscore Recall	Bertscore F1
0.6149	0.7163	500	0.6405	0.3775	0.1265	0.9079	0.9049	0.9064
0.4125	1.4327	1000	0.4899	0.2759	0.0851	0.9296	0.9283	0.9289
0.2366	2.1490	1500	0.4250	0.2544	0.0862	0.9404	0.9387	0.9395
0.233	2.8653	2000	0.4028	0.2305	0.0739	0.9438	0.9421	0.9429

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.1+cu128
• Datasets 4.4.2
• Tokenizers 0.22.1