whisper-large-v3-sandi-train-dev-4

This model is a fine-tuned version of ntnu-smil/whisper-large-v3-sandi-train-dev-1-merged on the ntnu-smil/sandi2025-ds dataset. It achieves the following results on the evaluation set:

• Loss: 0.9514
• Wer: 165.7353
• Cer: 153.9116
• Decode Runtime: 299.2678
• Wer Runtime: 0.1932
• Cer Runtime: 0.4197

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: 7e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 32
• total_train_batch_size: 1024
• optimizer: Use adamw_torch with betas=(0.9,0.98) and epsilon=1e-06 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• training_steps: 28

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Decode Runtime	Wer Runtime	Cer Runtime
2.4477	1.1435	7	1.2751	62.2053	233.9402	303.8935	0.1871	0.4849
1.1579	2.2870	14	1.1035	114.3075	211.3775	298.4955	0.1899	0.4722
1.0051	3.4305	21	0.9898	150.5656	182.6119	295.2723	0.1937	0.4415
0.9878	4.5740	28	0.9514	165.7353	153.9116	299.2678	0.1932	0.4197

Framework versions

• PEFT 0.15.1
• Transformers 4.48.3
• Pytorch 2.6.0
• Datasets 3.5.0
• Tokenizers 0.21.1