SsongSsong/model

whisper_ko_finetune100k

This model is a fine-tuned version of openai/whisper-base on the ksponspeech dataset. It achieves the following results on the evaluation set:

• Loss: 0.4177
• Cer: 15.0287

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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
0.7481	0.03	200	0.7555	44.8484
0.5393	0.06	400	0.5608	20.6969
0.5012	0.1	600	0.5293	17.3680
0.504	0.13	800	0.5075	17.5291
0.4934	0.16	1000	0.4927	15.6846
0.4958	0.19	1200	0.4838	17.7085
0.4697	0.22	1400	0.4768	18.5884
0.4501	0.26	1600	0.4688	16.5509
0.4505	0.29	1800	0.4617	17.2548
0.4434	0.32	2000	0.4566	15.7144
0.4482	0.35	2200	0.4515	16.0023
0.4333	0.38	2400	0.4465	14.8264
0.433	0.42	2600	0.4444	15.2950
0.4027	0.45	2800	0.4406	15.3715
0.4181	0.48	3000	0.4376	14.8904
0.4209	0.51	3200	0.4336	14.7853
0.4535	0.54	3400	0.4303	15.2584
0.4301	0.58	3600	0.4280	14.8824
0.4474	0.61	3800	0.4255	15.2287
0.4024	0.64	4000	0.4244	15.2515
0.3891	0.67	4200	0.4224	15.8675
0.411	0.7	4400	0.4206	15.0230
0.383	0.74	4600	0.4194	14.9293
0.4008	0.77	4800	0.4182	14.9818
0.4164	0.8	5000	0.4177	15.0287

Framework versions

• Transformers 4.36.0
• Pytorch 2.0.1+cu117
• Datasets 2.15.0
• Tokenizers 0.15.0