ntnu-smil/whisper-large-v3-turbo-sandi-train-dev-1-pure-transcript-32-2x-merged

whisper-large-v3-turbo-sandi-train-dev-1-pure-transcript-32-2x

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

• Loss: 0.6754
• Wer: 16.4068
• Cer: 11.7379
• Decode Runtime: 216.7200
• Wer Runtime: 0.1763
• Cer Runtime: 0.3328

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
• optimizer: Use OptimizerNames.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: 732

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer	Decode Runtime	Wer Runtime	Cer Runtime
0.841	0.1667	122	0.7927	18.6467	13.0967	208.7734	0.1720	0.3311
0.7312	1.0546	244	0.7441	17.7592	12.6026	214.5262	0.1676	0.3250
1.1313	1.2213	366	0.7133	17.1551	12.2452	213.0759	0.1802	0.3522
1.0806	2.1093	488	0.6927	16.7318	11.9823	212.2539	0.1729	0.3435
0.6408	2.2760	610	0.6810	16.5184	11.7819	212.9262	0.1715	0.3388
0.7177	3.1639	732	0.6754	16.4068	11.7379	216.7200	0.1763	0.3328

Framework versions

• PEFT 0.15.2
• Transformers 4.52.2
• Pytorch 2.8.0.dev20250319+cu128
• Datasets 3.6.0
• Tokenizers 0.21.1