seongsubae/openai-whisper-large-v3-turbo-ko-TEST

⚠️ Author's Note: This model was fine-tuned for sanity-checking purposes using only a single Korean dataset.
As a result, it may be overfitted and may not generalize well to other datasets.
You can find the training code and related resources in my GitHub repository: 2025-korean-asr-benchmark

openai/whisper-large-v3-turbo Korean - Fine-tuned

This model is a fine-tuned version of openai/whisper-large-v3-turbo on the Bingsu/zeroth-korean dataset. It achieves the following results on the evaluation set:

• Loss: 0.0733
• Wer: 4.3216

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: 128
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4622	0.1437	25	0.3690	19.2290
0.2	0.2874	50	0.1578	15.3290
0.1213	0.4310	75	0.1396	13.0703
0.1068	0.5747	100	0.1314	12.2572
0.1	0.7184	125	0.1242	11.0676
0.0922	0.8621	150	0.1181	10.6460
0.0895	1.0057	175	0.1122	9.6371
0.0667	1.1494	200	0.1098	9.2155
0.0608	1.2931	225	0.1049	8.4023
0.0608	1.4368	250	0.1007	7.6946
0.0577	1.5805	275	0.0992	7.4386
0.0591	1.7241	300	0.0953	6.5502
0.0547	1.8678	325	0.0920	5.9630
0.0518	2.0115	350	0.0885	5.5112
0.0299	2.1552	375	0.0878	5.8877
0.0311	2.2989	400	0.0872	4.8637
0.0319	2.4425	425	0.0895	5.2552
0.0363	2.5862	450	0.0869	5.1197
0.0325	2.7299	475	0.0851	4.9390
0.0331	2.8736	500	0.0849	4.7282
0.0314	3.0172	525	0.0805	4.9240
0.0196	3.1609	550	0.0805	4.5174
0.0164	3.3046	575	0.0820	5.4209
0.0166	3.4483	600	0.0807	6.1135
0.0153	3.5920	625	0.0775	3.9753
0.0127	3.7356	650	0.0741	4.8035
0.014	3.8793	675	0.0731	7.1827
0.012	4.0230	700	0.0719	4.9992
0.0067	4.1667	725	0.0744	4.5475
0.0061	4.3103	750	0.0732	5.2101
0.0053	4.4540	775	0.0736	4.4270
0.0061	4.5977	800	0.0743	4.8938
0.0048	4.7414	825	0.0740	5.3305
0.0045	4.8851	850	0.0733	4.3216

Framework versions

• Transformers 4.50.0.dev0
• Pytorch 2.4.0+cu121
• Datasets 2.21.0
• Tokenizers 0.21.1