Bateesa/whisper-small-jap

Whisper-medium-Multilingual

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

• Loss: 0.2195
• Wer: 14.1862

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: 64
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 5000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.4567	3.6364	1000	0.2337	28.9944
0.0403	7.2727	2000	0.2126	18.1510
0.0137	10.9091	3000	0.2142	17.1510
0.0013	14.5455	4000	0.2149	15.5248
0.0006	18.1818	5000	0.2195	14.1862

Framework versions

• Transformers 5.15.1
• Pytorch 2.13.0+cu130
• Datasets 2.21.0
• Tokenizers 0.22.2