Curiousfox/whisper_new_ver4

Hokkien-to-Tai Lo Whisper ver 4

This model is a fine-tuned version of openai/whisper-medium on the Common Voice 16.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4709
• Wer: 84.1740

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: 8
• eval_batch_size: 16
• 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: 1000
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3834	0.256	800	0.5102	83.2549
0.4023	0.512	1600	0.4974	82.0220
0.4024	0.768	2400	0.4877	83.6808
0.3819	1.024	3200	0.4820	80.9460
0.3125	1.28	4000	0.4772	79.4889
0.315	1.536	4800	0.4762	82.0220
0.3057	1.792	5600	0.4707	80.2062
0.3093	2.048	6400	0.4695	82.0444
0.2507	2.304	7200	0.4716	83.4342
0.2476	2.56	8000	0.4709	84.1740

Framework versions

• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0