huni0304/Whisper-small-koja

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

• Loss: 0.2226
• Wer: 17.5001

Model description

이 모델은 whisper-small을 한국어 및 일본어 데이터로 파인튜닝한 것이다.

Training and evaluation data

일본어 데이터 : mozilla-foundation/common_voice_11_0

한국어 데이터 : kresnik/zeroth_korean

총 한국어 데이터 : 22,263 발화

총 일본어 데이터 : 16,740 발화

Training data : 37,053 발화

Evaluation data : 1,950 발화

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.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
• training_steps: 6000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3482	0.4318	1000	0.3453	29.8426
0.289	0.8636	2000	0.2821	24.2086
0.1821	1.2953	3000	0.2548	20.7527
0.2005	1.7271	4000	0.2350	18.7315
0.135	2.1589	5000	0.2265	17.4421
0.1273	2.5907	6000	0.2226	17.5001

Framework versions

• Transformers 4.51.3
• Pytorch 2.7.0+cu126
• Datasets 2.16.0
• Tokenizers 0.21.1