Suchae/whisper-large-v3-ko-middlesenior-dialect-speech-v1.1

This model is a fine-tuned version of openai/whisper-large-v3 on the Suchae/whisper-large-v3-ko-middlesenior-dialect-speech-v1.1 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5249
• Cer: 14.4500

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-06
• train_batch_size: 10
• eval_batch_size: 5
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 80
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 50
• num_epochs: 1
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
1.333	0.0548	64	0.9131	19.4136
1.1099	0.1096	128	0.7550	17.2538
0.9577	0.1643	192	0.6955	17.2038
0.9198	0.2191	256	0.6615	15.8003
0.7995	0.2739	320	0.6357	16.5130
0.7898	0.3287	384	0.6150	15.8066
0.7344	0.3835	448	0.6022	14.9533
0.7035	0.4383	512	0.5846	14.3594
0.6936	0.4930	576	0.5711	16.6193
0.6427	0.5478	640	0.5602	14.7063
0.6365	0.6026	704	0.5530	15.0095
0.6107	0.6574	768	0.5440	14.5813
0.596	0.7122	832	0.5379	15.2315
0.5831	0.7670	896	0.5357	15.1377
0.5542	0.8217	960	0.5308	15.0314
0.5675	0.8765	1024	0.5277	15.2252
0.532	0.9313	1088	0.5252	15.6722
0.5255	0.9861	1152	0.5249	14.4500

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.1+cu118
• Datasets 3.0.0
• Tokenizers 0.19.1