mvbnh/whisper-medium-ko-new

Whisper Small Korea

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4319
• Wer: 26.9672

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: 16
• eval_batch_size: 8
• 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: 1500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0026	1.1111	50	0.3466	26.0886
0.0013	2.2222	100	0.3569	25.8212
0.0025	3.3333	150	0.3672	25.2865
0.003	4.4444	200	0.3578	25.5157
0.0091	5.5556	250	0.3599	25.3629
0.0072	6.6667	300	0.3682	25.9740
0.0054	7.7778	350	0.3785	26.8526
0.0093	8.8889	400	0.3764	27.1581
0.013	10.0	450	0.3886	28.5332
0.0146	11.1111	500	0.3900	27.5401
0.0128	12.2222	550	0.3917	27.3491
0.0054	13.3333	600	0.3926	26.5852
0.0029	14.4444	650	0.4281	28.4186
0.0062	15.5556	700	0.3957	27.5401
0.0062	16.6667	750	0.4080	27.7693
0.0023	17.7778	800	0.4151	27.4637
0.0034	18.8889	850	0.4153	28.2659
0.0009	20.0	900	0.4133	27.0053
0.0003	21.1111	950	0.4192	26.9672
0.0003	22.2222	1000	0.4223	26.9290
0.0002	23.3333	1050	0.4247	27.0053
0.0002	24.4444	1100	0.4266	26.9672
0.0002	25.5556	1150	0.4279	27.0817
0.0002	26.6667	1200	0.4290	27.0053
0.0002	27.7778	1250	0.4299	26.9672
0.0002	28.8889	1300	0.4306	26.9672
0.0002	30.0	1350	0.4312	27.0053
0.0002	31.1111	1400	0.4316	26.9672
0.0002	32.2222	1450	0.4318	26.9672
0.0002	33.3333	1500	0.4319	26.9672

Framework versions

• Transformers 4.42.4
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1