jethrowang/vanilla-whisper-base_hat

Whisper Base Hakka Condenser

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

• Loss: 0.1321
• Cer: 7.6600

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: 0.0001
• train_batch_size: 64
• eval_batch_size: 32
• seed: 42
• optimizer: Use 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: 488
• training_steps: 4880
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Cer
0.1299	0.9980	488	0.2820	17.9474
0.0685	1.9959	976	0.1976	12.9309
0.0314	2.9939	1464	0.1912	11.2525
0.0169	3.9918	1952	0.1744	13.9527
0.009	4.9898	2440	0.1621	9.9198
0.0027	5.9877	2928	0.1494	10.6526
0.0016	6.9857	3416	0.1451	8.7107
0.0006	7.9836	3904	0.1357	7.6889
0.0004	8.9816	4392	0.1329	7.6242
0.0001	9.9796	4880	0.1321	7.6600

Framework versions

• Transformers 4.49.0
• Pytorch 2.3.0
• Datasets 3.4.0
• Tokenizers 0.21.0