imTak/whisper_large_v3_turbo_korean_Develop

Whisper large v3 turbo Korean-Develop

This model is a fine-tuned version of imTak/whisper_large_v3_ko_ft_ft on the Develop dataset. It achieves the following results on the evaluation set:

• Loss: 0.3054
• Wer: 16.4370

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: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2119	1.9455	500	0.2721	22.6690
0.0714	3.8911	1000	0.2542	19.9135
0.0145	5.8366	1500	0.2417	18.5037
0.0018	7.7821	2000	0.2410	16.6453
0.0263	9.7276	2500	0.2818	19.4169
0.0179	11.6732	3000	0.2806	18.5838
0.008	13.6187	3500	0.2977	18.1032
0.0072	15.5642	4000	0.2920	17.8949
0.0011	17.5097	4500	0.2875	16.8376
0.0024	19.4553	5000	0.3072	17.8629
0.0009	21.4008	5500	0.2943	16.8536
0.0002	23.3463	6000	0.3041	16.8055
0.0001	25.2918	6500	0.2993	16.6773
0.0001	27.2374	7000	0.3016	16.4851
0.0001	29.1829	7500	0.3043	16.4050
0.0001	31.1284	8000	0.3054	16.4370

Framework versions

• Transformers 4.45.0
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3